55 plugins, 370 tool calls
Each plugin adds tool calls that an AI client, a pipeline or you can run. Open a plugin to see its tool calls.
A read tool changes nothing. A write tool changes something, such as sending a message or moving a file. 153 tools are read and 217 are write. The list is built from each plugin's manifest.yaml in the source.
Agents
Lets AI agents (Claude Code sessions, cloud agents, Claude on the phone) leave each other messages through Kamari. Each agent declares a name tied to its own credential and gets its own inbox; every step is audited.
8 tool calls: 3 read, 5 write
agent_closewriteMark a message sent to you as handled, so it leaves your open inbox. Requires message_id (from agent_inbox). Optional note (at most 500 characters). Closing twice is harmless. To answer, use agent_send with reply_to instead.
agent_directoryreadWho can I message? Lists the registered agents, what each one does, who owns it, when it last checked its inbox and how many open messages it has. No arguments. Call agent_register first.
agent_inboxreadCheck your messages. Lists messages sent to you, oldest first, without marking them read. Optional: unread_only (default true; false shows read and closed ones too), since (ISO 8601 time with offset; pass next_since to page on), limit (1-200, default 50; 1-20, default 20, with compact false) and compact (default true for short rows with an 80-character preview). Call agent_register first.
agent_managewriteMove one of your own agent handles to the credential you are calling with, or delete it. Requires action (rebind or remove) and handle. Use rebind when you reconnected and agent_register says the handle is registered to another credential: the handle and its inbox move to you. The old credential then holds no handle. A credential that already holds another handle must remove that one first. Needs the agents.manage permission.
agent_readwriteRead one message in full and mark it read. Requires message_id (msg_<epoch milliseconds>, from agent_inbox). Optional include_thread (true adds short rows for the rest of the conversation). You can read messages sent to you or by you.
agent_registerwriteTell Kamari which agent you are, once, before using the other agent_* tools. Requires handle (your address: 2-40 lower-case letters, digits and hyphens, e.g. vsc-kamari or claude-voice) and description (a few words, e.g. 'Claude Code session in the kamari repo'). The handle is bound to the credential you are calling with; calling again from the same credential only updates the description. A handle someone else holds is refused.
agent_sendwriteSend a message to another agent by name, for example the Claude Code session called vsc-kamari. Requires to (a handle from agent_directory), subject (at most 120 characters) and body (the message, at most 8000 characters by default). Optional: reply_to (the id of the message you are answering), needs_reply (true if you expect an answer) and refs ({quests, memory_keys, urls} lists). You are always the sender; call agent_register first.
agent_waitreadWait for new mail instead of checking again and again. Returns your unread messages as soon as one arrives (at once if some are already waiting), else an empty list after timeout_seconds (0-50, default 25). Optional since (ISO 8601 time with offset) skips messages created at or before it. Marks nothing read; use agent_read. Call agent_register first.
Ansible
Wraps the local ansible / ansible-playbook / ansible-inventory CLIs to run playbooks and ad-hoc commands from a sandboxed working directory.
5 tool calls: 1 read, 4 write
ansible_ad_hocwriteRun a one-off Ansible module against a set of hosts, without a playbook - e.g. restart a service or check disk space. Requires pattern (the host pattern to target, e.g. all, webservers, host1) and module (the module name, e.g. ping, command, shell, copy). Optional: module_args, working_dir, inventory, become=true for privilege escalation.
ansible_gather_factswriteCollect Ansible facts about hosts - OS, network interfaces, memory, disks - by running the setup module. Use it to answer questions about what a machine actually is. Requires pattern: the host pattern to target, e.g. all or webservers. Optional: working_dir, inventory.
ansible_list_inventoryreadList the fully resolved Ansible inventory as JSON, showing every group and host - use it to discover valid host patterns before running a playbook or ad-hoc command. All arguments are optional: working_dir and inventory override where the inventory is read from.
ansible_pingwriteCheck that hosts are reachable and that SSH authentication works, by running the Ansible ping module against them. Use it to debug before blaming a playbook. All arguments are optional: pattern selects the hosts (defaults to all), working_dir and inventory override where the inventory comes from.
ansible_run_playbookwriteRun an Ansible playbook against your inventory. Safe by default: it runs as --check (a dry run that changes nothing) unless you pass approve=true, so ask the user before setting that. Requires playbook: the playbook path relative to working_dir. Optional: working_dir, inventory, limit (host pattern), tags, extra_vars (key=value entries), var_file, approve.
Audio Stream Bridge
Browser mic/speaker bridge via WebSocket; routes TTS audio to browser clients
1 tool call: 1 read, 0 write
audio_bridge_statusreadReport the audio bridge's current state: whether it is running, how many bytes of audio are queued, and how many browser WebSocket clients are connected. Use to check whether TTS audio can actually reach a browser. Takes no arguments - call it bare.
AWS
Consolidated AWS service access — Polly TTS, SES email, SMTP email, and S3 file operations — all sharing one credential block.
7 tool calls: 2 read, 5 write
aws_polly_synthesizewriteTurn text into spoken audio using AWS Polly's cloud voices. Requires text: the words to speak. Optional: voice_id (a Polly VoiceId such as Joanna, Matthew or Brian), engine (neural, standard, long-form or generative), output_format (mp3, pcm or ogg_vorbis), polly_resume_task_id to resume a timed-out async job. Long text is synthesized asynchronously via S3.
aws_s3_downloadreadDownload one S3 object and return its content as text. Requires key: the object key (its path within the bucket), as shown by aws_s3_list. Optional bucket overrides the configured s3_bucket. Intended for text objects - binary content will not round-trip cleanly.
aws_s3_listreadList the objects in an S3 bucket, so you can discover the key to pass to aws_s3_download. All arguments are optional: bucket overrides the configured s3_bucket, prefix filters to keys starting with it (e.g. reports/), max_keys caps how many are returned (default 100).
aws_s3_uploadwriteUpload text content to an object in an S3 bucket. Requires key (the object key, i.e. its path within the bucket, e.g. reports/june.csv) and body (the UTF-8 text to store). Optional: bucket overrides the configured s3_bucket, content_type sets the MIME type (default text/plain).
aws_ses_sendwriteSend an email through AWS Simple Email Service. Requires to (the recipient's address) and subject (the subject line). Optional: body (plain text), html_body (sends multipart/alternative when given) and from_addr (overrides the ses_from_address config). Use aws_smtp_send instead for a non-AWS mail server.
aws_smtp_sendwriteSend an email through any SMTP server - not AWS-specific, despite living in the aws plugin. Requires to (the recipient's address) and subject (the subject line). Optional: body (plain text), html_body and from_addr (overrides the smtp_from_address config).
aws_transcribewriteTranscribe recorded speech to text using AWS Transcribe. Requires audio_b64: base64-encoded Int16 PCM audio (16kHz mono, as produced by the audio bridge) - not mp3 or wav. Optional language is a BCP-47 code such as en-US (default en-US).
Backup & Restore
Export the entire data directory (including the state.db — which holds the vault and every plugin's config) as a single zip, and restore it on a fresh install.
2 tool calls: 0 read, 2 write
backup_exportwriteExport the entire Kamari data directory as a single zip: state.db (which holds the vault and every plugin's config) plus every plugin's on-disk data. Use for 'back up Kamari' or before making a risky change. All arguments are optional - call it bare. The archive is written to .backup-out/ and its path returned, with content_b64 filled in only when the archive is small (<=32MB). If it reports the archive was too large to return inline, download it from /backup/export or feed the returned path straight to backup_restore.
backup_restorewriteRestore the whole Kamari data directory from a zip previously produced by backup_export. Destructive: it overwrites the current state.db, vault, plugin config and plugin data. Give exactly one of content_b64 (the base64-encoded zip) or path (an archive already on the server, under the data directory - use this for a large archive). Optional dry_run=true validates the archive and reports what would be restored without writing anything. An archive with no files, or one not written by backup_export (plugin.backup 0.4.0+; it carries a backup marker), is refused with an error before anything is touched - in dry_run too.
Blog
Author Markdown blog posts for an mkdocs site, published via a Bitbucket pull request per change.
6 tool calls: 3 read, 3 write
blog_create_articlewriteCreate a new Markdown blog article and open a Bitbucket pull request for it. Requires slug (filename stem - letters, digits, hyphens, underscores only), title, and content (Markdown body). Updates its local copy to the latest commit of the configured branch first, then commits the new file on a fresh branch and opens a PR - it never pushes directly to the configured branch.
blog_delete_articlewriteDelete an existing Markdown article and open a Bitbucket pull request for the removal. Requires path (from blog_list_articles). Updates its local copy to the latest commit of the configured branch first, then commits the deletion on a fresh branch and opens a PR - it never pushes directly to the configured branch.
blog_list_articlesreadList every Markdown article in the mkdocs content path of the configured docs repo, with title, file size and last-modified time. Takes no arguments - call it bare. Use the returned path with blog_update_article.
blog_pipeline_logreadRead the tail of one Bitbucket Pipelines step log for the docs site, e.g. to see why a build failed. Requires build_number (from blog_pipeline_status). Optional: step_name (default: the first failed step, else the last step) and max_bytes (1-65536, default 16384). Returns whole lines only, with credentials redacted. Read-only.
blog_pipeline_statusreadCheck whether a blog change built or deployed: lists recent Bitbucket Pipelines runs (newest first) with each run's state/result/stage and each step's state/result/stage. IMPORTANT: state IN_PROGRESS with stage HALTED or PAUSED means the run has stopped and is waiting for someone to act in Bitbucket - it is not running, so do not keep waiting. All arguments are optional individually, but give branch (e.g. the review branch from blog_create_article, or master) or commit_hash (full 40 characters), or both; limit is 1-10, default 5. Read-only. Use blog_pipeline_log to read a failed step's log.
blog_update_articlewriteUpdate an existing Markdown article and open a Bitbucket pull request for the change. Requires path (from blog_list_articles). Optional: title and content - omit either to leave it unchanged. Updates its local copy to the latest commit of the configured branch first, then commits the change on a fresh branch and opens a PR - it never pushes directly to the configured branch.
Calendar
Manage calendar events via a configurable provider — local JSON storage or Google Calendar.
6 tool calls: 3 read, 3 write
calendar_create_eventwriteAdd a new event to the calendar. Use for 'book', 'schedule' or 'put X in my diary'. Requires title (what the event is called) and start (an ISO date YYYY-MM-DD for an all-day event, or an ISO datetime like 2026-08-14T09:30:00 for a timed one). Optional: end (defaults to start), all_day, location, description, attendees, provider.
calendar_delete_eventwritePermanently remove an event from the calendar. Use for 'cancel my 3pm' or 'delete that meeting'. Requires event_id: the id from a calendar_list_events result. Look it up there first - deletion cannot be undone. Optional provider overrides the configured default backend.
calendar_get_eventreadFetch one calendar event's full details by id. Requires event_id: the id from a calendar_list_events result. Call calendar_list_events first to find it. Optional provider overrides the configured default calendar backend.
calendar_list_calendarsreadList the calendars the configured provider exposes (e.g. the separate Google calendars on an account), so you can tell which one events are being read from or written to. All arguments are optional: provider overrides the configured default backend.
calendar_list_eventsreadList calendar events, optionally limited to a date range - the way to answer 'what's on today', 'am I free on Friday' or 'what's coming up'. All arguments are optional: start_date and end_date are ISO dates (YYYY-MM-DD) bounding the range, and provider overrides the configured default calendar backend.
calendar_update_eventwriteChange an existing calendar event, updating only the fields you pass and leaving the rest alone. Use for 'move my 2pm to 3pm'. Requires event_id, from calendar_list_events. Every other argument is optional: title, start, end, all_day, location, description, attendees, provider.
Credential Administration
Admin-only MCP tools to mint, list, re-scope and revoke API keys and to register and edit OAuth login (OIDC) and service clients -- the same operations as the dashboard's admin pages, gated on the same auth:users permission.
10 tool calls: 0 read, 10 write
auth_admin_api_key_createwriteADMIN ONLY. Mint a new Kamari API key for an app or integration. Requires name (a label, e.g. the app), username (the existing Kamari user the key acts as) and project (the quest project it may read). Optional permissions narrows the key to a subset of that user's capabilities; omit to inherit the user's full set. The raw key is in the result once -- hand it to the app immediately; Kamari only keeps its hash. Caller needs auth:users.
auth_admin_api_key_listwriteADMIN ONLY. List API keys with their owner, project, capability subset and last use -- never the key itself. All arguments are optional: username filters to one owner. Use it to find a key_id for auth_admin_api_key_set_permissions or auth_admin_api_key_revoke. Caller needs auth:users.
auth_admin_api_key_revokewriteADMIN ONLY. Permanently revoke an API key so it stops working at once. Requires key_id (from auth_admin_api_key_list). Caller needs auth:users.
auth_admin_api_key_set_permissionswriteADMIN ONLY. Re-scope an existing API key without reissuing it. Requires key_id (from auth_admin_api_key_list) and permissions, the complete new capability list; [] makes the key inherit its owner's full set. Each capability must be held by the key's owner. Caller needs auth:users.
auth_admin_login_client_createwriteADMIN ONLY. Register an OIDC login client so another app can sign people in with Kamari. Requires client_name and redirect_uris (a LIST of every https callback URL). Optional: allowlist_users, backchannel_logout_uri, delegated_capabilities (tools a signed-in person may let the app use as them). The client_secret is in the result once -- hand it to the app immediately. Caller needs auth:users.
auth_admin_login_client_editwriteADMIN ONLY. Edit a login client. Requires client_id; then any of redirect_uris (the COMPLETE new list -- resend existing URLs to keep them), backchannel_logout_uri ('' clears) and delegated_capabilities ([] clears). All values are validated before anything is saved. Caller needs auth:users.
auth_admin_login_client_listwriteADMIN ONLY. List registered OIDC login clients (redirect URIs, allowlist, logout endpoint, delegated tools) -- never their secrets -- plus the capabilities a login client may be delegated. No required arguments. Use it to find a client_id for auth_admin_login_client_edit. Caller needs auth:users.
auth_admin_service_client_createwriteADMIN ONLY. Register an OAuth service client for a service that calls Kamari's /mcp from its own server with no person present (client_credentials). Requires client_name, scope_ceiling (capabilities it may ever get) and token_endpoint_auth_method ('client_secret_post' or 'private_key_jwt' with public_key or jwks_uri). Optional quest_projects. A minted client_secret is in the result once. Caller needs auth:users.
auth_admin_service_client_editwriteADMIN ONLY. Edit a service client. Requires client_id; then any of scope_ceiling, quest_projects, disabled, client_name, public_key / jwks_uri (replace the key) and token_endpoint_auth_method (switch secret <-> private_key_jwt; switching to a secret returns a new client_secret once). All values are validated before anything is saved. Caller needs auth:users.
auth_admin_service_client_listwriteADMIN ONLY. List OAuth service clients (ceiling, projects, auth method, key source, disabled flag) -- never their secrets -- plus the capabilities a service client may be granted. No required arguments. Use it to find a client_id for auth_admin_service_client_edit. Caller needs auth:users.
Credential Vault
Central encrypted credential store for provider secrets (AWS, SMTP, etc.)
2 tool calls: 1 read, 1 write
vault_getreadRead the stored credentials for one named provider out of Kamari's encrypted vault. Requires provider, the provider name to look up (e.g. aws, smtp, pushover). Returns found=false rather than an error when nothing is stored for that provider.
vault_setwriteStore or replace one provider's credentials in Kamari's encrypted vault, so plugins can reference them as {{vault.<provider>.<key>}}. Requires provider (the name to file them under, e.g. aws) and credentials (an object of key-value pairs). Optional merge=true keeps existing keys not present in this call; the default replaces the provider's entry outright.
Daily Plan
DailyPlan entity — the committed 1-3 tasks for today with the maxDailyTasks cap and human-in-the-loop approval gate
3 tool calls: 1 read, 2 write
dailyplan_approveTodaywriteRecord the user's explicit approval of a day's plan by stamping approvedAt on it - the human-in-the-loop gate. Call it only after the user confirms the plan that dailyplan_setToday committed. All arguments are optional: date is an ISO YYYY-MM-DD date, defaulting to today.
dailyplan_getTodayreadReturn a day's committed plan plus an isDayComplete flag showing whether every committed task is done. Use for 'what am I doing today' or 'is my day finished'. All arguments are optional: date is an ISO YYYY-MM-DD date, defaulting to today.
dailyplan_setTodaywriteCommit the short list of tasks for a day (hard cap of maxDailyTasks, default 3) and schedule each one via plugin.tasks. Use for 'plan my day' or 'these are today's tasks'. Requires taskIds: an array of task IDs to commit. Optional date: an ISO YYYY-MM-DD date, defaulting to today.
Device Tracker
A register of devices — each entry is a uniquely named device holding one interface (ethernet/wifi/bluetooth) per MAC, searchable by name or by any of its MACs in any notation
7 tool calls: 2 read, 5 write
device_tracker_add_macwriteGive an already-registered device another MAC address — its wifi or bluetooth interface alongside the ethernet one. Requires device_id (from device_tracker_list) and mac (any notation); kind is optional (ethernet, wifi, bluetooth, other). A MAC held by a different device is rejected, and re-adding one this device already has just relabels that interface.
device_tracker_createwriteRegister a new device in the device tracker. Use for 'remember this MAC' or 'add X to my devices'. Requires name (a friendly label, unique case-insensitively) and mac (its first interface's hardware address, written with or without ':' / '-' separators, in any case); kind is optional (ethernet, wifi, bluetooth or other). Add the device's other interfaces with device_tracker_add_mac rather than creating a second device.
device_tracker_deletewritePermanently remove a device from the register, which frees its name and all of its MAC addresses to be used again. Requires device_id, the id from a device_tracker_list result (format device<epoch>, e.g. device1786985825). Look the id up with device_tracker_list first — deletion cannot be undone.
device_tracker_getreadFetch one device's full details: every interface it holds with that interface's kind, plus when the device was registered and last changed. Requires device_id, the id from a device_tracker_list result (format device<epoch>, e.g. device1786985825). When you only know a name or a MAC, call device_tracker_list with search first to get that id.
device_tracker_listreadList the registered devices with all their interfaces, or look one up before getting, updating or deleting it. The only argument is optional: search, matched case-insensitively against the device name (substring) and against every MAC it holds, in any notation — 'AA:BB', 'aa-bb' and 'aabb' all find AA:BB:CC:DD:EE:FF. Omit it to list every device.
device_tracker_remove_macwriteRemove one interface from a device, e.g. after a network card is replaced. Requires device_id (from device_tracker_list) and mac, the address to drop, in any notation. The device's last remaining MAC cannot be removed — delete the whole device with device_tracker_delete instead.
device_tracker_updatewriteRename a registered device. Requires device_id (from device_tracker_list), field (only name is editable) and value (the new name; blank is rejected and it must not collide with another device). To change which MACs a device has, use device_tracker_add_mac and device_tracker_remove_mac instead.
Document Datastore
Local NoSQL document store (collections of schemaless JSON documents, key-value, field-filter queries). Provider-swappable for a real NoSQL backend.
7 tool calls: 4 read, 3 write
docstore_countreadCount how many documents a collection holds, without fetching any of them. Requires collection: the collection name, as listed by docstore_list_collections. Optional provider overrides the configured backend for this call.
docstore_deletewritePermanently delete one stored JSON document. Requires collection (the collection name) and doc_id (the document's key within it). Deletion cannot be undone, so read it with docstore_get first if you may need the contents.
docstore_getreadFetch one stored JSON document by its key. Requires collection (the collection name) and doc_id (the document's key within that collection). Use docstore_query when you only know field values rather than the key. Optional provider overrides the configured backend.
docstore_list_collectionsreadList the names of every collection in the document store, so you can discover the valid collection value for the other docstore tools. All arguments are optional: provider overrides the configured backend for this call.
docstore_migratewriteCopy every collection and document from one storage backend to another, e.g. when moving off the local JSON store. Requires from_provider (the backend to read from) and to_provider (the backend to write into). Both must be configured provider names.
docstore_putwriteStore a JSON document, creating the collection if needed and replacing any document already at that key. Requires collection (the collection name, e.g. notes), doc_id (the document's key within it) and document (the JSON object to store). The whole document is replaced, so send every field you want kept.
docstore_queryreadFind documents in a collection by matching field values - the way to search when you do not know a document's key. Requires collection (the collection name). Optional: filters, an object of field=value pairs that must all match (dot-notation works for nested fields), limit (default 100), provider.
Email
Send and receive email with swappable outbound and inbound providers
13 tool calls: 5 read, 8 write
email_config_createwriteCreate a named send/receive configuration -- required before email_send/email_receive/email_read etc can use it. Requires config_name (a unique label, e.g. 'gmail-invoices') and fields (the provider settings -- see email_config_list's docs for the field set per outbound_provider/inbound_provider). Optional allowed_principals lists other usernames/api-keys permitted to use this config (the caller is always added automatically); pass ['*'] to make it public. Use email_config_edit to change fields or the allow-list later.
email_config_deletewritePermanently delete a named configuration and stop any future email_send/email_receive from being able to use it (already-cached emails from it are not deleted). Requires config_name. Only callers already on the config's allow-list (or holding the bypass) may delete it.
email_config_editwriteChange fields and/or the allow-list of an existing named configuration. Requires config_name. Optional fields merges into the existing field set; optional allowed_principals, if given, REPLACES the entire allow-list (include the caller if they still need access). Only callers already on the config's allow-list (or holding the bypass) may edit it.
email_config_listreadList the named configurations the caller is permitted to use (by allow-list membership, or the pipelines.all-equivalent email.all bypass). Use this to find the config_name to pass to email_send/email_receive/email_read/etc.
email_countreadReturn how many emails a named configuration has cached in total and how many are unread, for a quick 'do I have new mail' check without listing them. Requires config_name (see email_config_list). Takes an optional unread_only flag, which is ignored - both totals are always returned.
email_fetchwriteFetch mail from a named IMAP configuration by time range, and optionally by sender or subject -- use for 'get today's mail' or 'get the latest mail from billing@example.com' instead of email_receive. Requires config_name (see email_config_list) and since (ISO 8601 date or date-time, at most 90 days ago). Optional: until (default now), from_addr (exact sender address), subject_contains, limit (default 50, max 200) and publish_events (default false). Returns emails[] newest first with each id, ready for email_get or for a pipeline tool's email_id. It never moves email_receive's watermark. With publish_events=false no pipeline fires, and a later email_receive will NOT publish those messages either, because it only publishes mail it newly caches.
email_getreadFetch one email in full, including its complete body, which email_read does not return. Requires config_name (the configuration that received it) and id: the email's ID exactly as returned in an email_read result. Use email_read first to find the id you want.
email_get_attachmentreadFetch metadata (and, for small files, base64 content) for one attachment stored from a received email. Requires config_name (the configuration that received it), message_key (the email's id/message_id) and attachment_id (the sha256 from that email's attachments[]). Returns a server-side path -- pass that to a downstream tool (e.g. pdf_extract_text) rather than the bytes, since large step payloads are dropped from pipeline context.
email_mark_readwriteMark one cached email as read, so it stops appearing in unread counts and unread_only listings. Requires config_name and id: the email's ID as returned by email_read. This only updates Kamari's local cache.
email_mark_spamwriteFlag one cached email as spam, or clear that flag again. Spam is hidden from email_read by default. Requires config_name and id (the email's ID from email_read). Optional spam defaults to true; pass spam=false to unmark a message wrongly flagged.
email_readreadList cached emails received by a named configuration - the normal way to answer 'check my email' or 'what's in my inbox'. This reads the cache only; call email_receive first to pull anything new. Requires config_name (see email_config_list). Other arguments are optional: limit (default 20), offset for paging, unread_only, spam_only, exclude_spam (default true).
email_receivewritePoll a named configuration's inbound provider for newly arrived emails and cache them locally so email_read and email_count can see them. Use before reading when the user expects something just-sent to be there. Requires config_name (see email_config_list).
email_sendwriteSend an email through a named outbound configuration. Use for 'email X' or 'send a mail about Y'. Requires config_name (see email_config_list), to (the recipient's address) and subject (the subject line). Optional: body (plain text), html_body (sends multipart when given), from_addr (defaults to the config's outbound_from_address) and attachments (array of {path, filename?, content_type?} -- server-side file paths, capped by config.email.max_attachment_bytes/max_attachments_per_message, same as inbound).
Exec
Runs a small, pre-programmed set of host diagnostic commands with no shell involvement and per-command permission gating -- never arbitrary shell/SSH access.
10 tool calls: 2 read, 8 write
exec_get_allowlistreadRead the argument character allowlist currently configured for one registered exec command. Requires command_id.
exec_get_domain_allowlistreadRead the hostnames exec_run_curl is currently permitted to reach. Takes no arguments. An empty list means every host is currently rejected.
exec_run_curlwriteRun curl against a caller-supplied https:// URL whose host is on the domain allowlist (see exec_set_domain_allowlist). Requires url. Optional: method (default GET), headers (object), body (string). Requires its own dedicated permission grant -- separate from every other exec command.
exec_run_data_disk_usagewriteSummarise disk usage per top-level entry under /data (one `du -sh` per child directory, capped at 50 entries). Takes no arguments. Requires its own dedicated permission grant.
exec_run_data_file_countwriteCount files under /data (runs `find /data -type f` and counts lines in Python -- no shell pipe). Takes no arguments. Requires its own dedicated permission grant.
exec_run_disk_usagewriteRun `df -h` on the host to show whole-disk usage. Takes no arguments. Requires its own dedicated permission grant -- granting this does not grant any other exec command.
exec_run_docker_logs_kamariwriteFetch the last N lines of Kamari's own stdout/stderr from its console log file under the data directory (what docker logs would show; no docker CLI is used). All arguments are optional: tail is a digits-only string (e.g. "200"), capped at 10000, defaulting to "200" if omitted. Requires its own dedicated permission grant -- separate from every other exec command.
exec_run_uptimewriteRun `uptime` on the host to show load averages and how long it has been running. Takes no arguments. Requires its own dedicated permission grant.
exec_set_allowlistwriteReplace the argument character allowlist for one registered exec command. Requires command_id. Optional: digits, lowercase, uppercase (booleans, all default false) and other_chars (list of single extra characters). Leaving everything false/empty rejects all arguments for that command.
exec_set_domain_allowlistwriteReplace the full set of hostnames exec_run_curl is permitted to reach. Requires domains (list of bare hostnames, no scheme/path/port). Replaces the entire list; an empty list rejects every host.
File Management
List, read, write, move, copy, and delete files hard-sandboxed to the plugin's own data root.
8 tool calls: 3 read, 5 write
files_copywriteCopy a file, or a whole directory tree, within Kamari's sandboxed file store. Requires src (the existing path, relative to the sandbox root) and dest (the path to copy it to). Use files_move instead to relocate rather than duplicate.
files_deletewritePermanently delete a file or directory from Kamari's sandboxed file store. Requires path, relative to the sandbox root. Optional recursive must be set true to delete a directory that is not empty; without it, a non-empty directory is refused rather than silently emptied.
files_listreadList the files and folders in a directory inside Kamari's sandboxed file store. Use it to browse before reading or writing, since every path is relative to the sandbox root and cannot escape it. All arguments are optional: path is a directory relative to that root, and empty (the default) lists the root itself.
files_mkdirwriteCreate a directory inside Kamari's sandboxed file store, including any missing parent directories along the way. Requires path: the directory path relative to the sandbox root, e.g. projects/notes. Succeeds quietly if the directory already exists.
files_movewriteMove or rename a file or directory within Kamari's sandboxed file store. Requires src (the existing path, relative to the sandbox root) and dest (the new path). Both must stay inside the sandbox. Use files_copy instead to keep the original.
files_readreadRead a text file's full contents out of Kamari's sandboxed file store. Requires path: the file's path relative to the sandbox root, e.g. notes/todo.txt. Use files_list first when you do not know the exact path. Binary files are not supported.
files_statreadGet metadata about a path - whether it exists, whether it is a file or directory, its size and modification time - without reading the contents. Requires path: the file or directory path relative to the sandbox root. Useful to check existence before reading or overwriting.
files_writewriteWrite text to a file in Kamari's sandboxed file store, creating any missing parent directories. Requires path (relative to the sandbox root) and content (the text to write). Optional append=true adds to the end of the file instead of replacing it - the default overwrites the file entirely.
Graph Datastore
Local graph store (nodes with labels/properties, directed typed edges, neighbour/traversal/path queries). Provider-swappable for a real graph DB backend.
10 tool calls: 5 read, 5 write
graph_add_edgewriteCreate a directed relationship from one node to another, or replace it if edge_id already exists. Requires from_id (the source node's id), to_id (the target node's id) and edge_type (the relationship label, e.g. KNOWS or OWNS). Optional: edge_id (auto-generated when omitted), properties, provider.
graph_add_nodewriteCreate a node in the graph, or replace it entirely if that id already exists. Requires node_id: a unique id you choose for it, e.g. person:alice. Optional: labels (types for the node, e.g. ["Person"]), properties (an object of attributes), provider. Connect nodes together with graph_add_edge.
graph_delete_edgewriteDelete one relationship between nodes, leaving both nodes themselves in place. Requires edge_id: the edge's id, as returned by graph_add_edge or shown on a graph_get_neighbors result. Use graph_delete_node to remove a node and all its edges instead.
graph_delete_nodewritePermanently delete a graph node and every edge connected to it, in or out. Requires node_id: the node's id, e.g. person:alice. This cascades to relationships, so anything linked through this node loses that connection.
graph_find_pathreadFind the shortest chain of relationships connecting two nodes, i.e. how they are related. Requires from_id (the starting node's id) and to_id (the node to reach). Optional: max_depth caps how many hops to search before giving up (default 10), provider overrides the backend.
graph_get_neighborsreadList the nodes directly connected to one node - one hop only. Use for 'who does X know' or 'what is attached to X'. Requires node_id, the node to start from. Optional: direction (out, in or both; default both), edge_type to only follow one relationship type, provider.
graph_get_nodereadFetch one graph node's labels and properties by id. Requires node_id: the node's id, e.g. person:alice. Use graph_match_nodes instead when you only know a label or property value rather than the exact id. Optional provider overrides the configured backend.
graph_match_nodesreadSearch for nodes by their label and/or property values, rather than by id - the way in when you do not know a node_id yet. All arguments are optional: label narrows to one node type, properties is an object of field=value filters that must all match, limit caps results (default 100).
graph_migratewriteCopy every node and edge from one graph backend to another, e.g. when moving off the local store. Requires from_provider (the backend to read from) and to_provider (the backend to write into). Both must be configured provider names.
graph_traversereadWalk the graph outward from a node, breadth-first, collecting everything reachable within a number of hops - the multi-hop version of graph_get_neighbors. Requires start_id: the node to start from. Optional: direction (out, in or both), max_depth (default 3), edge_type, provider.
Home Assistant
Subscribes to Home Assistant entity state changes over its websocket API and emits them as events pipelines can act on
6 tool calls: 2 read, 4 write
home_assistant_add_connectionwriteConnect Kamari to a Home Assistant instance and start streaming its entity state changes as events pipelines can trigger on. Requires name (a unique internal handle), host (e.g. homeassistant.local:8123) and access_token (a Home Assistant long-lived access token from Profile > Security). Optional: use_ssl (default true) and entity_ids to scope which entities stream - leaving it empty streams everything, which is a lot of traffic.
home_assistant_get_connectionreadFetch one Home Assistant connection including its stored access_token, for checking exactly what is configured. All arguments are optional, but pass either connection_id (from home_assistant_list_connections) or name (the internal handle, matched case-insensitively) or it cannot identify a connection.
home_assistant_list_connectionsreadList every registered Home Assistant connection with its live status and delivery stats, to find a connection_id or check whether events are actually arriving. Access tokens are withheld here - use home_assistant_get_connection to reveal one. All arguments are optional: search filters over name and host.
home_assistant_reconnectwriteForce a connection's websocket session to restart without changing any of its settings - the fix after resolving a network or firewall problem on the Home Assistant side. Requires connection_id: the id from home_assistant_list_connections.
home_assistant_remove_connectionwriteStop streaming from a Home Assistant instance and delete its connection entirely, so its entity events stop firing. Requires connection_id: the id from home_assistant_list_connections. Use home_assistant_update_connection instead if you only need to change a setting.
home_assistant_update_connectionwriteChange a Home Assistant connection's settings; changing host, access_token or use_ssl reconnects, while name and entity_ids apply live. Requires connection_id, from home_assistant_list_connections. Every other argument is optional and omitting one leaves it unchanged: name, host, access_token, use_ssl, entity_ids.
Host Metrics
Samples host CPU/memory/disk/load stats on a cron cadence and records them into plugin.metrics (service="host") so they can be queried, charted and alarmed on like any other metric.
1 tool call: 0 read, 1 write
host_metrics_samplewriteSample current host CPU/memory/disk/load stats and record them into plugin.metrics under service="host". Takes no arguments. Meant to be called periodically (e.g. by a cron pipeline rule), not ad hoc -- each call writes a fresh batch of metric points.
HTTP
Generic HTTPS-capable REST client for arbitrary method calls — used to send webhook requests and drive REST APIs from pipelines.
1 tool call: 0 read, 1 write
http_requestwriteSend an HTTP or HTTPS request to any REST API or webhook and return the response status, headers and body. Requires method (one of GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS) and url (must start with http:// or https://). Optional: headers, json_body (sent as application/json), body (raw text, ignored when json_body is set), timeout_seconds (default 30).
IMAP Mailbox
Generic IMAP/SMTP mailbox access — list folders, read, move, and send messages
4 tool calls: 1 read, 3 write
imap_list_foldersreadList the folders (mailboxes) available on the configured IMAP account, e.g. INBOX, Sent, Archive. Use it to discover valid folder names before calling imap_read or imap_move. Takes no arguments - call it bare.
imap_movewriteMove one email message into a different IMAP folder, e.g. to archive or file it. Requires uid (the message's IMAP UID, taken from an imap_read result) and dest_folder (the destination folder name, as listed by imap_list_folders). Optional source_folder is the folder the uid was read from (default INBOX); pass the same folder you gave imap_read.
imap_readwriteRead the most recent messages in an IMAP folder, returning each message's UID, sender, subject and body. Use for 'check my email' or 'what's in my inbox'. All arguments are optional: folder names the mailbox to read (default INBOX) and limit caps how many messages come back (default 20).
imap_sendwriteSend a plain-text email via the configured SMTP account. Use for 'email X about Y'. Requires to (the recipient's email address), subject (the subject line) and body (the plain-text message). HTML and attachments are not supported here.
Investec
Investec Open API banking -- accounts, balances, transactions, beneficiaries, cards (read-only, unverified -- see docs), webhook management (unverified, signature scheme unconfirmed -- see docs), and payments -- with a payment-safety pipeline (duplicate detection, rate limiting, and hold-for-approval above a configurable threshold).
21 tool calls: 12 read, 9 write
investec_approve_paymentwriteRelease a payment that was held for approval so it actually executes, after the user confirms it. Requires payment_id: the ID returned by investec_transfer or investec_pay_multiple when the payment was held, also listed by investec_list_pending_approvals. Duplicate and rate-limit checks are re-run before it goes out, then the account's current balance -- if there is not enough, the payment is parked as 'waiting for funds' and retried automatically every 6 hours instead of failing outright. Always replays against the environment the payment was originally held under; optional environment here only asserts it matches that value. Optional expected_amount (ZAR) is the amount the user was shown: pass it so that if the amount was edited since, nothing is sent and the call fails naming both amounts.
investec_delete_webhookwriteDelete an Investec-side webhook subscription. UNVERIFIED against a live account (quest1790454940) -- see the plugin docs. Requires webhook_id from investec_register_webhook or investec_list_webhooks. Does not delete the Kamari-side webhook.create endpoint that was receiving deliveries. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_accountsreadList the Investec bank accounts on the authenticated profile, with their account ids, names and types. Call this first - every other investec tool needs an account_id from here. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_balancereadGet one bank account's current and available balance. Use for 'how much money do I have'. Requires account_id: the Investec accountId from investec_get_accounts, not the human account number. Call investec_get_accounts first if you do not have it. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_beneficiariesreadList the saved payment beneficiaries on the profile, with the beneficiary_id each one needs for investec_pay_multiple. Call this before paying anyone to get the right id. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_card_transactionsreadList transactions for one card, optionally date-bounded. UNVERIFIED against a live account (quest1790454945) -- endpoint path and whether it keys on cardId vs accountId is a best-effort guess, see the plugin docs. Requires card_id from investec_list_cards. Optional from_date and to_date are inclusive ISO dates (YYYY-MM-DD). Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_pending_transactionsreadList transactions that have been authorised but have not yet posted to the account, so a balance can be reconciled against money already committed. Requires account_id: the Investec accountId from investec_get_accounts. Use investec_get_transactions for completed ones. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_proof_of_paymentreadGet a PDF proof-of-payment document for one posted transaction, returned as base64. UNVERIFIED against a live account (quest1790454693) -- confirm the response shape before relying on it in production. Requires account_id (the Investec accountId from investec_get_accounts) and transaction_id (from investec_get_transactions). Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_statementreadFetch a formal account statement PDF export for a date range (base64-encoded), for bookkeeping/reconciliation rather than the raw transaction list investec_get_transactions returns. Requires account_id (the Investec accountId from investec_get_accounts), from_date and to_date (both required, inclusive ISO dates YYYY-MM-DD, to_date must not be before from_date). UNVERIFIED against a live Investec account -- see the plugin docs. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_get_transactionsreadList the posted (completed) transactions on a bank account, optionally limited to a date range. Use for 'what did I spend' or 'show my statement'. Requires account_id, the Investec accountId from investec_get_accounts. Optional from_date and to_date are inclusive ISO dates (YYYY-MM-DD). Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_list_cardsreadList cards on the authenticated profile. UNVERIFIED against a live account (quest1790454945) -- endpoint path/response shape is a best-effort guess, see the plugin docs. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_list_payment_historyreadList past decisions on held payments (approved or rejected), newest-decided first, with amount/beneficiary/note/outcome/decided_at. Use it to answer 'what happened to this payment' once it is no longer in investec_list_pending_approvals. Optional environment (sandbox|production) filters to decisions recorded under that one environment.
investec_list_pending_approvalsreadList the payments currently held for approval and waiting on a decision, with the payment_id each one needs. Use it to answer 'is anything waiting on me' before approving or rejecting. Optional environment (sandbox|production) filters to payments held under that one environment.
investec_list_webhooksreadList Investec-side webhook subscriptions on this profile. UNVERIFIED against a live account (quest1790454940) -- see the plugin docs. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_pay_multiplewritePay one or more saved beneficiaries from a bank account - real money leaving the account, so confirm details with the user first. Requires account_id (the source accountId from investec_get_accounts) and payments (an array of {beneficiary_id, amount, my_reference, their_reference} -- all four required, with beneficiary_id from investec_get_beneficiaries; a blank/missing my_reference or their_reference is rejected, quest1790461502). Optional note (max 500 chars) explains the payment for whoever approves it on /web/payments if it gets held. Optional source_email ({from_addr, subject, body_text, recipient_name}) attaches the verbatim triggering email for held-payment review; recipient_name (max 100 chars, no newlines or control characters) makes the confirmation email greet the payee as 'Hi <name>,' instead of 'Hi,'. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox -- the payment-safety pipeline (duplicate/rate-limit/hold-for-approval) applies the same either way. Payments over the configured threshold are held for approval.
investec_recheck_waiting_for_fundswriteRe-check balance for every payment parked as 'waiting for funds' and execute any that now has enough. Runs automatically every 6 hours via a self-registered pipeline rule -- normally there is no need to call this by hand. The payment rate limit applies: a payment it blocks stays waiting for funds for a later run. Optional environment (sandbox|production) filters this run to payments held under that one environment; each waiting payment always replays against its own recorded environment regardless.
investec_register_webhookwriteRegister an Investec-side webhook so Investec pushes events (e.g. transactions) to a URL instead of polling. UNVERIFIED against a live account (quest1790454940), and signature verification is NOT wired up -- see the plugin docs before relying on this in production. Requires url: call webhook_create (plugin.webhook) first with a name like 'investec-transactions', then pass its returned url_path (as a full https URL) here. Optional webhook_type defaults to 'transaction'. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox.
investec_reject_paymentwriteDiscard a payment that is being held for approval, so it is never executed. Requires payment_id: the ID returned by investec_transfer or investec_pay_multiple when the payment was held, also listed by investec_list_pending_approvals. This cannot be undone - the payment must be recreated. Refused for a payment in status 'executing' (being sent, or stuck): an admin resolves a stuck one with investec_resolve_stuck_payment. Optional environment only asserts it matches the environment the payment was held under.
investec_resolve_stuck_paymentwriteAdmin only: record whether a payment stuck in 'executing' actually left the bank, after checking the bank. Requires payment_id (from investec_list_pending_approvals), verdict ('sent' or 'not_sent') and reason (what you checked, 1-500 characters). sent marks the payment executed; not_sent puts it back to held for approval, so it must be approved again. Never contacts the bank. Needs auth:users plus the investec.resolve_stuck_payment permission. Refused while an approve or recheck is still sending the payment, and for any payment that is not executing.
investec_transferwriteMove money between two of the user's own Investec accounts. Requires from_account_id (source accountId), to_account_id (destination accountId) - both from investec_get_accounts - amount (a number in ZAR), my_reference and their_reference (statement references; a blank/missing value is rejected, quest1790461502). Optional note (max 500 chars) explains the payment for whoever approves it on /web/payments if it gets held. Optional source_email ({from_addr, subject, body_text, recipient_name}) attaches the verbatim triggering email for held-payment review; recipient_name (max 100 chars, no newlines or control characters) makes the confirmation email greet the payee as 'Hi <name>,' instead of 'Hi,'. Optional environment (sandbox|production) overrides config.investec.environment for this call only, default sandbox -- the payment-safety pipeline (duplicate/rate-limit/hold-for-approval) applies the same either way. Large transfers are held for approval rather than executed immediately.
investec_update_held_amountwriteChange the amount of a payment that is held for approval, before it is approved -- for example when the invoice pipeline computed the wrong figure. Requires payment_id (from investec_list_pending_approvals) and new_amount (ZAR, greater than 0, at most 2 decimal places); optional reason is kept in the payment's change history. Needs the investec.update_held_amount permission, which approve rights do not include. Refused once the payment is waiting for funds, being executed, approved or rejected, for multi-line payments, and above config.investec.max_held_amount_increase_percent over the original amount. Never executes the payment: it stays held until investec_approve_payment.
Local LLM (Gemma)
CPU-friendly local LLM for basic conversation, backed by Ollama
1 tool call: 0 read, 1 write
llm_chatwriteSend a message to the local Ollama-hosted LLM and get a conversational reply. Use for open-ended chat, rewriting or summarising text when no other tool fits. When tool-calling is enabled the model may only call tools that both the plugin-llm user and you hold, and only effect: read tools when config.llm.tools_read_only is on. Requires message (the text to send to the model). Optional: system (system-prompt override) and model (Ollama model name override).
Logs
Read Kamari's own logs over MCP -- bounded, redacted queries plus a per-caller "what broke since I last looked" summary.
3 tool calls: 3 read, 0 write
logs_queryreadQuery Kamari's own logs (not remote syslog). All arguments are optional: filter by since/level/plugin_id/contains, bounded+capped limit, newest-first. Every field is redacted before being returned. Call bare for the most recent entries across all plugins.
logs_retention_statusreadShow the log retention in force (retention_days) and the age of the oldest retained record in the log ring and in the audit log. Takes no arguments.
logs_summarize_sincereadErrors/warnings since this caller's last look (or an explicit cursor), grouped by normalised signature with counts. All arguments are optional. Advances the caller's cursor unless dry_run=true. Every field is redacted before being returned.
Memory
LLM/user persistent and temporary memory items with per-item read access control.
8 tool calls: 2 read, 6 write
memory_create_persistentwriteSave a fact for the long term so it survives across conversations - use for anything the user says to remember, or preferences worth keeping. Requires content: the text to remember. Optional: key (a short descriptive label) and readable_by (usernames besides the owner who may read it). Use memory_create_temp instead when the fact should expire.
memory_create_tempwriteSave a short-lived fact that auto-expires, for things only relevant for a while (a code, a running total, today's context). Requires content: the text to remember. Optional: ttl_seconds (lifetime, defaults to config.memory.default_temp_ttl_seconds), key (short label), readable_by (other usernames allowed to read it).
memory_deletewritePermanently delete one stored memory item. Use when the user says to forget something. Requires memory_id, the id from memory_list (format mem_<epoch>, e.g. mem_1786561481). Only the item's owner or a holder of the memory.all permission may delete it.
memory_getreadFetch one stored memory item's full content by its id. Requires memory_id, the id from memory_list (format mem_<epoch>, e.g. mem_1786561481). Only the owner, a username in the item's readable_by list, or a holder of the memory.all permission can read it.
memory_listreadList the memory items the calling user can see - their own, ones shared with them, or all of them with the memory.all permission. Use this to find an item's memory_id, fetch named items in one call, or answer 'what do you remember'. All arguments are optional: kind (persistent or temp), key_prefix (array, e.g. ["kamari:"]), keys (array of exact keys; unmatched ones come back in missing), updated_since (ISO 8601 with offset; matches items created, updated, shared or unshared since then), compact (true for short previews, 50 per page), limit (1-200) and offset. Called bare it returns every item in full, which can be very large: prefer compact=true. Pages report total, has_more and next_offset. Expired temp items are never returned.
memory_sharewriteGrant other users read access to one of your memory items. Requires memory_id (the item's id from memory_list, format mem_<epoch>) and usernames (an array of usernames to grant). Only the owner or a memory.all holder can share; use memory_unshare to take access away again.
memory_unsharewriteRevoke other users' read access to one of your memory items. Requires memory_id (the item's id from memory_list, format mem_<epoch>) and usernames (an array of usernames to revoke). Only the owner or a memory.all holder can do this; the owner always keeps access.
memory_updatewriteChange one stored memory item's content and/or key in place; use this instead of delete + recreate, so the item keeps its id, owner and created_at. Every argument is optional on its own, but a call needs exactly one of memory_id (format mem_<epoch>) or key (exact match; refused with the matching ids if several items you may change share it), plus content (not blank), new_key, or both. Only the owner or a memory.all holder may update; any other item reads as not found. Renaming onto a key the owner already uses is refused.
Metrics
Domain-specific metrics layer (record/query/alarm/rollup) built on top of plugin.tsdb, with a local catalog of services/metric names for filtering, an alarm engine that publishes events pipelines can act on, and a cron-triggerable rollup capability that compacts old raw points into summary-stat buckets to bound storage growth.
11 tool calls: 6 read, 5 write
metrics_aggregatereadDownsample a service's measurements into fixed-width time buckets, giving one value per bucket - the right tool for charting or 'what was the average X'. Requires service, from metrics_list_services. Optional: metric_name, type, from_ts, to_ts, fn (count, sum, min, max or avg; default avg), bucket_seconds (default 60).
metrics_alarm_deletewriteDelete the alarm configured for one metric, so it stops evaluating and can no longer fire. Requires service and metric_name, the pair identifying the metric, as listed by metrics_alarm_list. Use metrics_alarm_set with enabled=false to pause an alarm instead of removing it.
metrics_alarm_getreadFetch the alarm configured for one metric, including its threshold rule and whether it is currently breaching. Requires service and metric_name, the pair identifying the metric, as listed by metrics_list_metrics. Returns nothing configured if that metric has no alarm.
metrics_alarm_listreadList every configured alarm across all services along with its current breaching or ok state - use it to answer 'is anything alarming right now'. Takes no arguments - call it bare.
metrics_alarm_setwriteCreate or update an alarm that fires when a metric breaches a threshold on X out of the last Y points. Requires service, metric_name, comparison (one of >, >=, <, <=, ==, !=), threshold (the number compared against), x_count (how many breaching points trigger it) and y_count (the window size in points). Optional: missing_policy (breaching, ok or ignore), enabled.
metrics_list_metricsreadList the known service/metric_name/type combinations, i.e. which metrics actually exist and can be queried, charted or alarmed on. All arguments are optional: service scopes the list to one service (omit for every service).
metrics_list_servicesreadList every service that has ever recorded a metric, so you can discover the valid service value for the other metrics tools. Start here when you do not already know the service name. Takes no arguments - call it bare.
metrics_queryreadList the raw recorded measurement points for a service - use it to inspect actual values rather than a summary. Requires service, as listed by metrics_list_services. Optional: metric_name and type to narrow it (see metrics_list_metrics), from_ts and to_ts as inclusive ISO or epoch bounds, limit (default 1000).
metrics_recordwriteRecord a single measurement against a service, and evaluate any alarm configured for it. Requires service (what the metric belongs to, e.g. kamari-api), metric_name (e.g. request_latency or cpu_usage) and value (the number). Optional: type (units such as seconds or percentage), ts (ISO or epoch, defaults to now), tags.
metrics_record_batchwriteRecord many measurements in one call, evaluating each affected metric's alarm once. Requires points, a list of objects each with service, metric_name and value, and optionally type, ts and tags. Prefer this over repeated metrics_record calls: the whole batch is one write. Points missing a service, metric_name or numeric value are skipped and reported in the skipped count.
metrics_rollupwriteCompact a metric's raw points older than a configurable age into fixed-width time buckets of summary stats (min, max, avg, count, sum, p95, p99), deleting the raw points once compacted by default. Intended to run on a schedule via a pipelines cron trigger. Requires service and metric_name. Optional: type, older_than_seconds (default 604800 = 7 days), bucket_seconds (default 3600 = 1 hour), destructive (default true; set false to keep raw points alongside the new rollup stats).
Movies
Manage your movie library via a configurable provider (Radarr) — search and add movies, check whether a movie is downloaded, see upcoming release dates, the full library, download history, the download queue and wanted/missing movies.
9 tool calls: 8 read, 1 write
movies_addwriteAdd a movie to the library so the provider (Radarr) will download it. Use for 'download/get me the movie X' - run movies_search first to get the id. Requires tmdb_id: the TMDb id from a movies_search result. Optional: quality_profile_id, root_folder, monitored, minimum_availability (announced, inCinemas or released), search_for_movie, provider.
movies_calendarreadList movies with release dates coming up, so you can answer 'what films are out soon'. All arguments are optional: days sets how far ahead to look (default 30), include_unmonitored=true also returns unmonitored films, provider overrides config.movies.provider.
movies_historyreadList recent movie download activity (grabbed, imported, failed, deleted), newest first, one page at a time. Use for 'what movies downloaded recently' or 'did X fail'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.movies.provider.
movies_history_sincereadList every movie download event since a given date, newest first. Use for 'what movies came in since Monday'. Requires date: an ISO 8601 date or date-time such as 2026-10-01 or 2026-10-01T08:30:00Z (no timezone means UTC; must not be in the future). Optional provider overrides config.movies.provider.
movies_libraryreadList every movie in the library with its download status and release dates, sorted by title. Use for 'what movies do I have' or 'list my films'. All arguments are optional: provider overrides config.movies.provider.
movies_queuereadShow what movies are downloading right now, with percent complete, size, time left and any error. Use for 'what's downloading' or 'how long until X finishes'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.movies.provider.
movies_searchreadSearch the movie provider (Radarr) for films matching a title, returning candidates you can then add with movies_add. Use for 'find the movie X'. Requires term: the movie title or search text. Optional provider overrides config.movies.provider for this one call.
movies_statusreadCheck whether a movie is already downloaded and report its cinema, digital and physical release dates. Use for 'do I have X yet' or 'when is X out'. All arguments are optional, but pass either tmdb_id (preferred when known) or term (the title to match in the library) or you will not identify a film; provider overrides config.movies.provider.
movies_wanted_missingreadList monitored movies that have no downloaded file yet (Radarr's Wanted > Missing), one page at a time. Use for 'which movies are still missing'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.movies.provider.
Network Tools
Network diagnostics and lookup utilities — MAC address analysis, ICMP/TCP reachability checks, iterative DNS resolution with RDAP registrar lookup
4 tool calls: 4 read, 0 write
net_tools_dns_lookupreadResolve a domain iteratively from the root servers (root to TLD to authoritative, following referrals rather than the OS resolver) and pull registration details from the TLD's RDAP service. Use for 'who owns X', 'what are X's nameservers' or to debug DNS. Requires name: the domain to resolve. Optional: record_type (A, AAAA, NS, CNAME, MX, TXT, SOA, PTR; default A), include_registrar_info, timeout, max_referrals.
net_tools_icmp_pingreadPing a host with ICMP echo requests and report per-packet round-trip time, packet loss and min/avg/max latency. Use for 'is X up' or 'how's the latency to X'. Requires host: the hostname or IP to ping. Optional: count (echo requests to send, default 4) and timeout (seconds for the whole run, default 10).
net_tools_mac_lookupreadAnalyse a MAC address: report whether it is a randomised (locally administered) address and look up its vendor via any enabled info providers. Requires mac, the address in any common format such as AA:BB:CC:DD:EE:FF, AA-BB-CC-DD-EE-FF or AABBCCDDEEFF.
net_tools_tcp_pingreadMeasure TCP connect latency to a host and port, and estimate the hop count via a TTL ramp - use this when ICMP is filtered and net_tools_icmp_ping gets no reply, or to check a specific service's port is open. Requires host (hostname or IP) and port (the TCP port). Optional: attempts, timeout, max_hops.
Notifications
Ordered per-user notification queue, with per-item dismiss permission (the recipient, or only the author).
4 tool calls: 2 read, 2 write
notification_createwriteQueue an in-app Kamari notification for a specific user; it stays in their list until dismissed. Use to hand a message or action item to a person - for a phone push use pushover_send instead. Requires recipient (the target username, e.g. spac3) and message (the text to show).
notification_dismisswriteDismiss (clear) a single notification so it stops appearing in the recipient's list. Requires notification_id, the id string from notification_list or notification_list_sent (format notif_<epoch>, e.g. notif_1786561481). The recipient can only dismiss it when it was created with dismissible_by=user; the author always can.
notification_listreadList the calling user's own undismissed notifications, highest priority first. Use to answer 'do I have any notifications' or 'is anything waiting for me'. All arguments are optional: pass include_dismissed=true to also return ones already cleared.
notification_list_sentreadList notifications the calling user has authored, for any recipient, including whether each was dismissed yet. Use to check what you already sent someone before sending a duplicate. Takes no arguments - call it bare.
NZB Downloads
Manage NZB (Usenet) downloads via a configurable provider (SABnzbd) — add downloads by URL, watch the queue, and review history.
4 tool calls: 2 read, 2 write
nzb_addwriteQueue a Usenet download by handing the provider (SABnzbd) an NZB file URL. Requires url: a link to the .nzb file. Optional: name overrides the display name, category picks a download category (empty uses the configured default), provider overrides config.nzb.provider for this call.
nzb_historyreadList finished Usenet downloads, both completed and failed, so you can confirm something landed or find why it failed. All arguments are optional: limit caps how many history entries come back (default 20) and provider overrides config.nzb.provider.
nzb_queuereadList what is downloading right now, with per-item progress, speed and time remaining. Use for 'how's my download going'. All arguments are optional: limit caps the entries returned (0, the default, means all) and provider overrides config.nzb.provider.
nzb_removewriteCancel a queued download or delete a history entry. Requires nzo_id: the provider's download id, taken from an nzb_queue or nzb_history result. Optional from_history=true targets the history list instead of the active queue; provider overrides config.nzb.provider.
Obsidian
Read/write notes and properties in a pre-configured Obsidian vault folder, with instant file/folder change notifications.
14 tool calls: 5 read, 9 write
obsidian_list_subscriptionsreadList the vault paths currently being watched for changes, so you can see what will raise obsidian.file_changed events, or find the exact path string to pass to obsidian_unsubscribe. Takes no arguments - call it bare.
obsidian_read_filereadRead one Obsidian note, returning both its body text and its YAML frontmatter properties. Use for 'what's in my note about X'. Requires path: the note's vault-relative path including the .md extension, e.g. Projects/Todo.md.
obsidian_read_propertiesreadRead only a note's YAML frontmatter properties - tags, status, dates and any other metadata - without pulling back the whole body. Requires path: the note's vault-relative path, e.g. Projects/Todo.md. Use obsidian_read_file when you also need the content.
obsidian_subscribewriteWatch a note or folder in the vault so that any change publishes an obsidian.file_changed event that pipelines can trigger on. All arguments are optional: path is a vault-relative file or folder and an empty string (the default) watches the entire vault; recursive decides whether changes in subfolders count (default true).
obsidian_sync_configwriteSet how syncing behaves for the linked vault. All arguments are optional and anything left out falls back to the matching config.obsidian.sync_* setting: mode (bidirectional, pull-only or mirror-remote), conflict_strategy (merge or conflict), excluded_folders, included_extensions.
obsidian_sync_list_remotereadList the Obsidian Sync vaults available to the signed-in account, so you can pick the right value for config.obsidian.sync_vault_name before running obsidian_sync_setup. Requires that obsidian_sync_login has already succeeded. Takes no arguments - call it bare.
obsidian_sync_loginwriteSign the obsidian-headless CLI in to Obsidian Sync, using the email and password already stored in config.obsidian. Run it once before the other sync tools will work. All arguments are optional: mfa_code carries the current two-factor code, which must be passed per call because it is one-time and is never stored.
obsidian_sync_logoutwriteSign out of Obsidian Sync and clear the credentials the obsidian-headless CLI has stored on this machine. Use it to disconnect the device or before logging in as a different account. Takes no arguments - call it bare.
obsidian_sync_nowwriteRun a single one-off sync between the local vault folder and Obsidian Sync, then stop. Use it to pull changes on demand or push edits immediately; it works even when continuous syncing is switched off in config. Takes no arguments - call it bare.
obsidian_sync_setupwriteLink the local vault folder to the remote Obsidian Sync vault, registering this machine as a sync device. It uses config.obsidian.vault_root, sync_vault_name and sync_device_name, so set those first - obsidian_sync_list_remote shows the valid vault names. Takes no arguments - call it bare.
obsidian_sync_statusreadReport whether the background continuous-sync daemon is currently running, along with its recent output - the tool for 'is my vault actually syncing' or diagnosing why changes are not appearing. Takes no arguments - call it bare.
obsidian_unsubscribewriteStop watching a path that was previously registered with obsidian_subscribe, so its changes no longer raise events. All arguments are optional: path is the vault-relative path to stop watching, and an empty string refers to the whole-vault subscription. Check obsidian_list_subscriptions for the exact path.
obsidian_update_propertieswriteMerge key/value pairs into a note's YAML frontmatter, leaving the body completely untouched - the safe way to tag or restatus a note. Requires path (the note's vault-relative path) and properties (an object of keys to set). Existing keys with the same name are overwritten; others are left alone.
obsidian_write_filewriteWrite an Obsidian note's body, creating the note and any missing parent folders. Requires path (vault-relative, e.g. Projects/Todo.md) and content (the body text below the frontmatter). This replaces the whole body, so read it first if you mean to append. Existing frontmatter is preserved; optional properties merge into it, and create=false refuses to make a new file.
PDF Reader
Extracts text and structured fields from text-layer PDFs, with optional positional (bbox) reads and an opt-in OCR fallback for scanned pages
3 tool calls: 3 read, 0 write
pdf_extract_fieldsreadExtract named fields from a PDF by applying a regex pattern (one capture group) per field, optionally scoped to a page/bbox, casting to number or string. This is the call an invoice-processing pipeline uses to pull totals and deductions. Set ocr:true to fall back to OCR (via plugin.vision) instead of raising when a field's page has no text layer.
pdf_extract_textreadExtract text from a PDF. All arguments are optional: pass path or content_b64 to identify the PDF, and optionally page (1-based) and bbox ({x0, top, x1, bottom}, or set relative:true for 0-1 fractions) to restrict the read to one region of one page. Set ocr:true to fall back to OCR (via plugin.vision) instead of raising when a page has no text layer.
pdf_inforeadReport page count, encryption state, metadata and per-page sizes for a PDF. All arguments are optional: pass path (resolved within the plugin's data dir) or content_b64 to identify the PDF.
People
A general-purpose person-info store — each person holds multiple optional fields (phone, email, company, domain, beneficiary_id, account_id, surname, display_name), stored as JSON files
5 tool calls: 2 read, 3 write
people_createwriteAdd a new person to the store. Use for 'save this contact' or 'add X as a person'. Requires name: the person's name, kept exactly as given. Everything else is optional: phone, email, company, domain (the company's email domain, which lets other plugins correlate people), beneficiary_id and account_id (Investec payment-routing ids for paying this person), surname, and display_name (what emails greet the person with; greeting rule is display_name if set, else name).
people_deletewritePermanently delete a person from the store. Requires person_id, the id from a people_search result (format person<epoch>, e.g. person1786561481). Look the id up with people_search first - deletion cannot be undone.
people_getreadFetch one person's full details. Requires person_id, the id from a people_search result (format person<epoch>, e.g. person1786561481). When you only know the person's name, call people_search with search first to get that id.
people_searchreadSearch/list people. All arguments are optional. With no arguments, lists everyone. Pass `search` for a case-insensitive substring match across name and every field (may return several loosely-matching people). Pass BOTH `field` (one of name, phone, email, company, domain, beneficiary_id, account_id, surname, display_name) and `value` for an exact, single-field match instead — e.g. field='email', value='someone@example.com' — which returns only exact matches so a caller can confidently take result[0]. Use it to look someone up, or to find a person_id before getting, updating or deleting a person.
people_updatewriteChange one single field on an existing person. Requires person_id (from people_search), field (exactly one of name, phone, email, company, domain, beneficiary_id, account_id, surname, display_name) and value (the new text; pass an empty string to clear an optional field). Call it once per field you need to change.
Pipelines
Pipeline automation — listens to all events and routes them through configurable multi-step pipelines
16 tool calls: 5 read, 11 write
pipelines_add_rulewriteCreate an automation that runs a sequence of steps whenever a matching event fires - the way to set up 'when X happens, do Y'. Requires name (a human-readable label), trigger (the event name to listen for, an fnmatch glob such as sms.*) and actions (the ordered array of steps, each an object with a type plus that type's fields). Optional: conditions to further filter the event payload, enabled (default true). A step may read an earlier invoke_tool step's result as {{_step<index>_<key>}} (or an earlier wait_for_event step's as _wait<index>_<key>); steps number from 0, a loop body numbers its own steps from 0, and a reference that can never resolve (missing step, wrong step kind, or a step at or after the one reading it) is refused. Each pipeline is also the tool pipeline_<name> (lowercased, punctuation as _); a name giving a tool name another pipeline or tool already uses is refused.
pipelines_debug_continuewriteResume a paused pipeline debug session until the next step is a breakpoint or the pipeline ends, then return the full state. Always runs at least one step. Requires session_id, from pipelines_debug_start. Optional: breakpoints, a replacement list of step paths (same format as pipelines_debug_start).
pipelines_debug_inspectreadRead a pipeline debug session's full state without running anything: every context variable (trigger fields, vars.*, _step<N>_* results, the current loop item), step records so far, the next step, loop position and captured messages. Requires session_id, from pipelines_debug_start.
pipelines_debug_list_sessionsreadList your own pipeline debug sessions that have not expired yet - paused ones you can still step, and recently ended ones you can still inspect. Takes no arguments - call it bare.
pipelines_debug_startwriteOpen a step-through debug session for a pipeline, to watch it run one step at a time and inspect every variable at each pause. Starts paused before the first step with nothing executed, and works on a disabled pipeline. Requires rule_id (the pipeline's UUID, from pipelines_list_rules). Optional: payload (the trigger event fields the steps start from) and breakpoints (0-based step paths such as "2", or "2.1" for step 1 inside the loop at step 2). Steps later run for real as the pipeline's owner - an invoke_tool step really calls its tool. Returns the session state including session_id.
pipelines_debug_stepwriteRun exactly one step of a paused pipeline debug session, then pause again and return the full state (context variables, step results, next step, captured log lines and events). Stepping onto a loop enters its body with the first item bound. Requires session_id, from pipelines_debug_start.
pipelines_debug_stopwriteEnd a paused pipeline debug session without running its remaining steps. If any step already ran, the partial run is saved to the pipeline's run history (outcome stopped). Requires session_id, from pipelines_debug_start.
pipelines_disable_rulewriteSwitch an automation off without deleting it, so it stops firing but keeps its configuration for later. Use this rather than pipelines_remove_rule when pausing something temporarily. Requires rule_id: the pipeline's UUID, from pipelines_list_rules.
pipelines_edit_rulewriteChange an existing automation, updating only the fields you pass and leaving the rest as they are. Requires rule_id, the pipeline's UUID from pipelines_list_rules. Every other argument is optional: name, trigger, actions (replaces the whole step list), conditions, enabled. Ownership is unchanged - use pipelines_set_owner for that. Submitted actions or conditions holding a _step<index>_/_wait<index>_ reference that can never resolve are refused, as in pipelines_add_rule, and so is a new name whose pipeline_<name> tool name another pipeline or tool already uses.
pipelines_enable_rulewriteTurn a disabled automation back on so it starts responding to its trigger event again. Requires rule_id: the pipeline's UUID, from pipelines_list_rules. Enabling an already-enabled pipeline is harmless.
pipelines_list_active_runsreadList the pipeline runs currently in flight across every pipeline, with a concurrency count and a stale marker on runs that look stuck. Read-only: it never cancels or closes a run. Takes no arguments - call it bare.
pipelines_list_rulesreadList the automation pipelines the caller can see, with each one's trigger, conditions, steps and rule_id. Start here to find the rule_id needed by the edit, enable, disable, remove and run-history tools, or to answer 'what automations do I have'. Takes no arguments - call it bare.
pipelines_list_runsreadList a pipeline's recent executions, newest first (up to 50), showing each run's steps and whether it succeeded - the tool for debugging 'why didn't my automation work'. Requires pipeline_id: the pipeline's UUID, from pipelines_list_rules.
pipelines_remove_rulewritePermanently delete an automation pipeline, so it stops firing and its configuration is gone. Requires rule_id: the pipeline's UUID, from pipelines_list_rules. Use pipelines_disable_rule instead to stop it temporarily while keeping the configuration.
pipelines_rerun_runwriteReplay a past run of a pipeline using that run's exact original trigger payload, as a brand-new run with a fresh run_id (never reusing the old run's ids). Use this to retry a run that failed because of a rule bug, once the rule has been fixed, without waiting for a matching real-world event. Requires pipeline_id (the pipeline's UUID, from pipelines_list_rules) and run_id (the past run's run_id, from pipelines_list_runs).
pipelines_set_ownerwriteReassign an automation to a different owner. This matters for more than bookkeeping: the owner is the permission boundary that decides what the pipeline is allowed to do at runtime. Requires rule_id (the pipeline's UUID) and new_owner (the username taking it over). Requires the pipelines.all permission.
Plugin Scaffold
Generates a Kamari plugin skeleton (manifest.yaml, plugin.py, __init__.py) from a manifest YAML string. The manifest is the source of truth.
1 tool call: 0 read, 1 write
plugin_scaffoldwriteGenerate a complete Kamari plugin skeleton on disk from a manifest. Requires manifest_yaml: the full manifest.yaml content as one string, following the schema spelled out in this capability's description (api_version, kind, id, name, version, description, spec.capabilities, spec.mcp_tools, spec.config_schema). Writes manifest.yaml, plugin.py (one method per capability), __init__.py and a tests/ skeleton into plugins_root/<shortname>/. Optional: plugins_root (defaults to the built-in plugins package), dry_run=true to preview which files would be written without touching disk, overwrite=true to replace existing files. The manifest is validated first and no files are written if validation fails.
Price Tracker
Barcode -> South African retailer price tracker. Resolves a barcode to a product name, searches a configurable list of SA retailer domains for that product, and records a dated price history per site so a re-scan shows price movement over time.
21 tool calls: 6 read, 15 write
pricetracker_add_domainwriteAdd a retailer domain to a product tracked by name. Requires product (its key, from pricetracker_list_tracked_products) and domain (e.g. 'takealot.com'). Adding to a product that followed the global default list gives it its own list.
pricetracker_add_listingwritePin a known listing URL for a product so refreshes always fetch it, useful when discovery misses it. Requires product (key) and url (must be on one of the product's domains).
pricetracker_delete_barcode_entrywriteDelete a single barcode's entry from the local barcode database, after which that barcode no longer resolves to a product name. Requires barcode: the EAN or UPC to remove. Use pricetracker_set_barcode_entry instead if the name is merely wrong.
pricetracker_delete_barcode_sourcewriteDelete every barcode entry that came from one dataset, plus that import's run record - the clean-slate step before re-importing a source. Requires source: the dataset name, as shown by pricetracker_list_barcode_db_sources (e.g. openfoodfacts). Manually added entries are not touched.
pricetracker_getreadReturn the prices already stored for a barcode without touching the network - the fast read for 'what did we find for this product'. Requires barcode: the product's EAN or UPC, as listed by pricetracker_list_products. Optional: sort_by (price or date, default price) and order (asc or desc, default asc).
pricetracker_get_barcode_entryreadLook up what product name a barcode resolves to in the local barcode database, along with which source supplied it and when. Requires barcode: the EAN or UPC to look up. Use pricetracker_set_barcode_entry to correct a wrong or missing name.
pricetracker_get_productreadRead what is stored for a product tracked by name, with no network calls: cheapest in-stock listing per domain, all listings, bundles flagged separately, gone listings, price history and the last diagnostics per domain. Requires product (key). Optional history_limit.
pricetracker_import_barcode_dbwriteBulk-import barcode-to-product-name rows from an open dataset into the local barcode database, so barcodes resolve to names without a live lookup. Requires source: one of openfoodfacts, openproductsfacts, openbeautyfacts, openpetfoodfacts, openfoodfacts_dump (the Open Food Facts bulk file, downloaded once and filtered locally, which avoids the API rate limit) or wikidata (items with a GTIN, CC0). Each source stores its licence and attribution, shown by list_barcode_db_sources. Rows are filtered by config.pricetracker.barcode_db_country_tag; the run is long and can be cancelled mid-way. A partial import resumes from the next page on the next call; pass full_reimport true to start over.
pricetracker_list_barcode_db_sourcesreadList the barcode datasets that have been imported, with each one's row count, last-imported time and status - use it to check whether an import finished or is still running. Takes no arguments - call it bare.
pricetracker_list_productsreadList every product currently being tracked, with its name, cheapest known price and when it was last checked. Use it to find a barcode for the other tools, or to answer 'what am I tracking'. Takes no arguments - call it bare.
pricetracker_list_tracked_productsreadList products tracked by name with their domains, active listing count, cheapest in-stock listing and last refresh time. Takes no arguments - call it bare.
pricetracker_refreshwriteRe-search the retailers for a barcode already being tracked and append today's prices to its history, so you can see whether anything moved. Requires barcode: the product's EAN or UPC. Unlike pricetracker_get this makes live network calls; unlike pricetracker_track it always re-searches.
pricetracker_refresh_allwriteRefresh every product tracked by name; this is what the daily scheduled rule calls. Progress is saved per product, so a run cut short by a restart continues where it stopped when called again (restart=true begins afresh). Known pages are re-fetched every run; discovery repeats weekly. Optional restart. Slow, makes live network calls.
pricetracker_refresh_productwriteCrawl now: find and re-check a product's listings on every domain, record prices and history, mark vanished listings gone, and return per-domain diagnostics (including bot blocking). Requires product (key). Makes live network calls and is slow because of per-domain rate limits.
pricetracker_removewriteStop tracking a product and delete its whole stored price history. Requires barcode: the product's EAN or UPC, as listed by pricetracker_list_products. The history cannot be recovered afterwards - re-tracking starts from scratch.
pricetracker_remove_domainwriteStop tracking one retailer domain for a product; past listings and history stay as a record. Requires product (key) and domain. The last domain cannot be removed - use pricetracker_remove_product.
pricetracker_remove_productwriteStop tracking a product by name and delete its listings, history and diagnostics; cannot be undone. Requires product (key).
pricetracker_seed_historywriteAdd dated prices noted by hand to a product's history for one domain. Requires product (key), domain and entries: a list of {recorded_at: '2026-10-01', price: 74999, currency?, url?, in_stock?, is_bundle?}. Repeating the same entries adds nothing.
pricetracker_set_barcode_entrywriteManually set the product name a barcode resolves to, for correcting a bad entry or adding one the imported datasets missed. Requires barcode (the EAN or UPC) and product_name (the name to record). The entry is stored with source 'manual', so a re-import will not overwrite it.
pricetracker_trackwriteStart tracking a product's price: resolve its barcode to a product name, search the configured South African retailer sites, and record what each one charges. Use for 'what does this cost' or 'track this product'. Requires barcode: the product's EAN or UPC number. Optional product_name_hint skips the barcode lookup and searches by that name instead.
pricetracker_track_productwriteStart tracking a product by name across retailer domains; the plugin finds the listings itself on every refresh. Requires name (e.g. 'Hisense L9Q'). Optional: domains (e.g. ['hirschs.co.za']; omit to use the global default list), aliases (other names shops use), ean, refresh_now (also run the first crawl now). Without an EAN the key is a slug such as hisense-l9q. Tracking the same key again merges aliases and domains.
Pushover Bot
Send push notifications via the Pushover service
1 tool call: 0 read, 1 write
pushover_sendwriteSend a push notification to the user's phone or desktop via Pushover. Use whenever asked to alert, ping, text or 'notify me' about something. Requires message (the notification body text). Optional: title, priority (lowest/low/normal/high/emergency), sound, url, url_title, user_key; retry and expire apply only when priority=emergency.
Quest Manager
Creates and manages quests stored as JSON files
19 tool calls: 7 read, 12 write
quest_add_updatewriteAppend a timestamped progress note to a quest's trail - the normal way to record what you did without editing the original description. Requires quest_id and message (the update text). Optional file_name plus file_content_b64 attach a file to the note. Also requires identity: name yourself (model plus session or agent name, 1-128 characters) so the update history shows which agent wrote the entry.
quest_checklist_addwriteAdd one to-do item to a quest's checklist, for tracking sub-steps within the quest. Requires quest_id and text (what needs actioning). Optional actor (human, agent or orchestrator; default agent): add actor=human, with how_to (one line of exact steps), for every hand test or owner action you report, then move the quest to Needs human. Optional due_by (ISO-8601 date). The new item's zero-based index is returned; use it with quest_checklist_set_done or quest_checklist_remove.
quest_checklist_removewriteDelete one item from a quest's checklist entirely - use quest_checklist_set_done instead if the item is simply finished. Requires quest_id and index, the item's zero-based position as shown in the checklist from quest_get. Removing shifts the indexes of later items.
quest_checklist_set_donewriteTick or untick one checklist item on a quest. Requires quest_id, index (the item's zero-based position, as shown in the checklist from quest_get) and done (true to mark it complete, false to reopen it).
quest_createwriteOpen a new quest (work item / ticket) to track one piece of work. Requires subject (a one-line title), description (the full detail of what needs doing) and type (an integer priority; by default 1 Urgent, 2 High, 3 Normal, 4 Low, 5 Backlog, but labels are user-configurable, so call quest_list_types for the live list). Optional: project to group it, and file_name plus file_content_b64 to attach a file. Also requires identity: name yourself (model plus session or agent name, 1-128 characters) so the update history shows which agent wrote the entry.
quest_getreadFetch one quest in full: description, status, owner, update history, checklist and attachments. Requires quest_id, the id from quest_list (format quest<epoch>, e.g. quest1786561481). Optional changes_limit (newest N trail entries of any kind, 0 for none; changes_total gives the full count) and include_description (false omits it) shrink the reply for a brief read. A brief read is only safe when human_summary_is_empty is false: when it is true, read the full trail. Only the quest's owner, or a caller holding the quest.all permission, can read it.
quest_get_attachmentreadDownload a file previously attached to a quest, returned base64-encoded. Requires quest_id and path, the attachment's stored path exactly as listed in the attachments array of quest_get. Call quest_get first to discover valid paths.
quest_listreadList quests the caller can see, to find an existing one before opening a new quest or to review what is outstanding. All arguments are optional: status defaults to 'open' (New + WIP + Needs human + Blocked + Paused) and also accepts New, WIP, Needs human, Blocked, Paused, Closed, Cancelled or all; actor='human' returns only quests with an open human checklist item, in ANY status including Closed and Cancelled, plus items (each open human item with quest, text, how_to and age); view='for_me' is that listing newest first; project narrows to one project; query free-text searches id, subject, description and project; sort='queue' puts queued quests first in queue order, then the rest by priority and age. Every row carries queue_position (1-based, or null if not queued).
quest_list_projectsreadList the distinct project names that have quests, each with a count, so you can discover the valid value for the project argument of quest_list or quest_create. All arguments are optional: status scopes the count (defaults to all).
quest_list_typesreadList the integer priorities valid for the type argument of quest_create and quest_update_quest, each with its user-configured label (e.g. 1 = Urgent), plus default_type to use when unsure. Takes no arguments.
quest_list_viewsreadList the built-in quest views (For me, Queue, Priority) with their definitions. For me is every open checklist item with actor human on a quest of ANY status; use quest_list view=for_me to fetch it. Takes no arguments.
quest_queue_getreadRead a project's processing queue: the order quests are WORKED in, which is separate from priority (quest_list sorts, the queue orders). Requires project. Returns version (pass it back as expected_version), quest_ids head first, and compact rows (quest_id, subject, type, status, position). Quests behind the head are processed after it; Paused and Blocked quests keep their place and are skipped until they are workable again.
quest_queue_movewriteMove one quest in a project's processing queue, or add it. Requires project, quest_id, identity (name yourself) and exactly one of position (1-based, 1 = head), before (a queued quest id) or after (a queued quest id). Optional expected_version refuses the move if the queue changed since you read it.
quest_queue_removewriteTake one quest out of a project's processing queue; it then follows by priority and age. Requires project, quest_id and identity (name yourself). Optional expected_version refuses the change if the queue changed since you read it.
quest_queue_setwriteReplace a project's whole processing queue. Requires project, quest_ids (the complete new order, head first, each id once, each an open quest of that project), expected_version (from quest_queue_get; 0 when there is no queue yet) and identity (name yourself). A stale expected_version is refused with the current version. Quests that finish (Closed or Cancelled) leave the queue on their own.
quest_set_ownerwriteReassign a quest to a different owner, which is what controls who can see it. Requires quest_id and new_owner (the username to hand it to). Only a caller holding the quest.all permission may do this.
quest_update_human_summarywriteSet a quest's human_summary: a short agent-written recap of what the quest is about and where it currently stands, so the next agent can orient without re-reading everything. Requires quest_id and value (the summary text, which replaces any previous one).
quest_update_questwriteEdit one human-authored field on an existing quest. Requires quest_id, field (exactly one of subject, description, project, type or model_tier) and value (replacement text that overwrites the field entirely; for type an integer priority from quest_list_types; for model_tier haiku, sonnet or opus, or empty to clear). Optional note records why (for model_tier, the routing reason). Use quest_add_update instead to append progress without rewriting the description.
quest_update_statuswriteMove a quest through its lifecycle: New to WIP to Closed or Cancelled, with Needs human (your work is done but a person must do or test something; needs an open human checklist item), Blocked (stalled externally) and Paused (deliberately parked) all reachable from New or WIP. Requires quest_id and new_status (one of New, WIP, Needs human, Blocked, Paused, Closed, Cancelled). Closing additionally requires commit_hash, the git hash of the commit that did the work - commit before you close - and is refused while a human checklist item is open, unless waive_human_items=true with a note. Optional note records why. Also requires identity: name yourself (model plus session or agent name, 1-128 characters) so the update history shows which agent wrote the entry.
Reflections
ReflectionLog entity — persist and retrieve daily/weekly review logs
2 tool calls: 1 read, 1 write
reflections_getReflectionsreadRetrieve recent reflection log entries, newest first. Use for 'what did I reflect on last week' or to review past daily or weekly summaries. All arguments are optional: type filters to daily or weekly (omit for both), limit caps how many entries come back (default 10).
reflections_logReflectionwriteSave a daily or weekly reflection entry to the reflection log. Use after a day-end or week-end review, or for 'log my reflection'. Requires type (either daily or weekly) and summary (the narrative text of the review). Optional: completedTaskIds, carriedTaskIds, droppedTaskIds, moodOrEnergy, date.
Review Page Verification Cases
MCP tools for the verification ledger behind the Review page (save/list/edit/run/delete per-capability test cases) -- wraps the same core VerificationLedger and gating rules as the /api/verification/* HTTP routes.
6 tool calls: 1 read, 5 write
review_add_casewriteSave a reusable Review-page test case for one plugin capability. Requires plugin_id, capability, name (unique within the plugin, no '.' or '{}') and payload (arguments matching the capability's input_schema; a value may reference an earlier case as '{{case name.field}}'). Optional expect_contains (substring asserted against the JSON result), expect_error (true for a case meant to be refused: it passes only when the call fails with an error containing expect_contains) and safe_for_replay (joins 'Run all safe tests' bulk replay; defaults off the capability's effect).
review_edit_casewriteUpdate an existing Review-page test case's name, payload and/or expect_contains in place. Requires plugin_id, capability and case_id; name/payload/expect_contains/expect_error are each optional -- omit a field to leave it unchanged. Changing the payload (or name, or expect_contains) resets the case's verdict to untested.
review_list_casesreadList the saved test cases for one plugin capability, each with its last verdict and run metadata. Requires plugin_id and capability.
review_preview_runwriteReally invoke a capability against an in-progress payload that has not been saved as a case yet, so a caller can see the real result shape before choosing a case name or expect_contains. Requires plugin_id, capability and payload. If the capability declares effect: write, the caller must also pass confirm=true AND hold the verification.trigger_write permission -- same gate as review_run_case. Never touches the verification ledger: nothing is saved, recorded, or judged.
review_remove_casewriteDelete one saved Review-page test case. Requires plugin_id, capability and case_id (as returned by review_add_case or review_list_cases).
review_run_casewriteReally invoke a saved case and record its verdict, exactly like the Review page's Run button. Requires plugin_id, capability and case_id. If the capability declares effect: write, the caller must also pass confirm=true AND hold the verification.trigger_write permission -- this tool never auto-runs a mutating case.
Samba File Access
List, read, and write files on configured network-share servers
3 tool calls: 2 read, 1 write
samba_listreadList the files and folders on a configured network share (Samba/SMB). Use it to browse a share before reading or writing a file. Requires server: the share name exactly as configured in config.samba.servers. Optional path, which defaults to that server's configured default_path.
samba_readreadRead one file's contents off a configured network share. Requires server (the share name from config.samba.servers) and path (that file's path on the share). Call samba_list first when you do not already know the exact path.
samba_writewriteCreate or overwrite a file on a configured network share. Requires server (the share name from config.samba.servers), path (the destination file path on that share) and content (the full text to write). An existing file at that path is overwritten without warning.
Shortlinks
Create, read and delete l.spac3.co.za redirect links stored in the link_payload DynamoDB table.
3 tool calls: 1 read, 2 write
shortlinks_create_linkwriteCreate an l.spac3.co.za short link. Requires slug (lowercase letters, digits, single hyphens; max 64) and target (https URL). Fails if the slug already exists or is reserved - it never overwrites. Returns the public link_url.
shortlinks_delete_linkwriteDelete an l.spac3.co.za redirect link by slug. Only pure redirect links can be deleted; page, webhook and metadata entries are refused. CloudFront may keep serving the old redirect briefly.
shortlinks_get_linkreadLook up where an l.spac3.co.za short link points. Requires slug. Only redirect links are returned; page and webhook entries are refused.
Shows
Manage your TV show library via a configurable provider (Sonarr) — search and add series, check whether an episode is downloaded, see upcoming air dates, download history (overall, since a date, per show), the download queue, and wanted missing / cutoff-unmet episodes.
10 tool calls: 9 read, 1 write
shows_addwriteAdd a TV series to the library so the provider (Sonarr) will download it. Use for 'download/follow the show X' - run shows_search first to get the id. Requires tvdb_id: the TheTVDB id from a shows_search result. Optional: quality_profile_id, root_folder, monitored, season_folder, search_for_missing, provider.
shows_calendarreadList TV episodes airing soon, so you can answer 'what's on this week'. All arguments are optional: days sets how far ahead to look (default 7), include_unmonitored=true also returns unmonitored series, provider overrides config.shows.provider.
shows_episodereadLook up a series' episodes and report which are downloaded and when each airs. Use for 'has the latest episode of X landed' or 'when is S02E03 out'. All arguments are optional, but pass either series_id (preferred when known) or term (the series title) to identify the show; season and episode narrow it further (0 means all).
shows_historyreadList recent TV download activity across all series (grabbed, imported, failed, deleted), newest first, one page at a time. Use for 'what episodes downloaded recently'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.shows.provider.
shows_history_seriesreadList the download history of one TV show, newest first. Use for 'what happened with X' or 'why is X S02 not downloading'. All arguments are optional, but pass either series_id (preferred when known) or term (the series title) to identify the show; season narrows it to one season (0 means all), provider overrides config.shows.provider.
shows_history_sincereadList every TV episode download event since a given date, newest first. Use for 'what episodes came in since Monday'. Requires date: an ISO 8601 date or date-time such as 2026-10-01 or 2026-10-01T08:30:00Z (no timezone means UTC; must not be in the future). Optional provider overrides config.shows.provider.
shows_queuereadShow which TV episodes are downloading right now, with percent complete, size, time left and any error. Use for 'what's downloading' or 'how long until X finishes'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.shows.provider.
shows_searchreadSearch the TV provider (Sonarr) for series matching a title, returning candidates you can then add with shows_add. Use for 'find the show X'. Requires term: the series title or search text. Optional provider overrides config.shows.provider for this one call.
shows_wanted_cutoffreadList monitored TV episodes whose downloaded file is below the quality profile cutoff and still wants an upgrade (Sonarr's Wanted > Cutoff Unmet), with each file's current quality. Use for 'which episodes need a better copy'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.shows.provider.
shows_wanted_missingreadList monitored TV episodes that have aired but have no downloaded file (Sonarr's Wanted > Missing), one page at a time. Use for 'which episodes am I missing'. All arguments are optional: page (default 1), page_size (1 to 250, default 20), provider overrides config.shows.provider.
SMS
Send and receive SMS messages via a configurable provider (MikroTik RouterOS native API)
7 tool calls: 3 read, 4 write
sms_deletewriteDelete a message from the provider's device inbox, e.g. to free space on a full SIM. Requires message_id: the provider id from the '.id' field of an sms_read result (e.g. *1). This does not touch Kamari's local copy - use sms_delete_local for that.
sms_delete_localwriteDelete one SMS from Kamari's local store. Requires id: the message id in <+E164>_<timestamp> form, as returned by sms_list. This is independent of sms_delete, which removes a message from the provider's device inbox instead.
sms_downloadwritePull the provider inbox and durably store any messages Kamari has not recorded yet, so they survive being cleared off the device. Safe to re-run - already-stored messages are skipped. All arguments are optional: limit caps how many inbox messages are scanned (default 100).
sms_getreadFetch one locally stored SMS in full. Requires id: the message id in <+E164>_<timestamp> form, exactly as returned by sms_list (e.g. +27821234567_1786561481). Call sms_list first to find the id.
sms_listreadList the SMS messages Kamari has stored locally, newest first - use this rather than sms_read for history, since the device inbox gets cleared. All arguments are optional: limit caps how many come back (default 50) and offset pages through older ones.
sms_readreadRead messages straight from the provider's SIM/modem inbox. This is the live device inbox, not Kamari's own store - use sms_list for messages already downloaded. All arguments are optional: limit caps how many messages are returned (default 20).
sms_sendwriteSend an SMS text message through the configured provider (MikroTik RouterOS). Use for 'text X' or 'send an SMS'. Requires phone_number (the recipient's number in international form, e.g. +27821234567) and message (the text, max 160 characters - anything longer is truncated).
Speech to Text
Transcribes audio segments to text
1 tool call: 0 read, 1 write
stt_transcribewriteTranscribe recorded speech audio into text. Use when given a voice note or audio clip that needs turning into words. Requires audio_b64: the audio data, base64-encoded. Optional: language, a BCP-47 code such as en-US (default en-US).
Syslog
Receives syslog messages (UDP/TCP, RFC 3164/5424) and routes each one through configurable regex rule groups keyed by sender IP — actions are listen (emit event), log (store to disk), or ignore (drop).
5 tool calls: 2 read, 3 write
syslog_add_rule_groupwriteCreate a syslog rule group: match incoming messages by sender IP, then apply an ordered list of message-text regex rules to decide what happens to each. Requires name (a label), default_action (listen, log or ignore - used when no rule matches) and rules (the ordered rule array, first match wins). Optional match_pattern is a regex on the source IP, defaulting to .* (any sender).
syslog_edit_rule_groupwriteChange an existing syslog rule group; only the fields you supply are altered, the rest are left as they are. Requires group_id, the id from syslog_list_rule_groups. Optional: name, match_pattern, default_action (listen, log, ignore) and rules, which replaces the whole ordered rule list when given.
syslog_list_rule_groupsreadList every configured syslog rule group with its source-IP pattern, default action and ordered rules. Use it to find a group_id before editing or removing a group, or to see why a message was handled a certain way. Takes no arguments - call it bare.
syslog_query_messagesreadSearch the stored syslog messages, e.g. to see what a device logged on a given day. All arguments are optional: date is a YYYY-MM-DD day defaulting to today UTC, source_ip filters to one exact sender IP, and limit caps how many messages return (default 100). Only messages whose matched action was 'log' are stored.
syslog_remove_rule_groupwritePermanently delete a syslog rule group, so its sender IPs fall back to whatever other group matches them. Requires group_id, the id of the group to remove as shown by syslog_list_rule_groups. Look the id up there first - this cannot be undone.
Tasks
Task entity — capture/clarify/triage/do/stale lifecycle, MCP tools only
12 tool calls: 4 read, 8 write
tasks_capturewriteCapture a new to-do into the inbox verbatim, with no structuring needed - the fast path for 'remind me to X' or 'add a task'. Requires title: the raw text of the item, kept forever as rawCapture. Optional source records where it came from, e.g. email, stt, manual. Clarify it later with tasks_clarifyTask.
tasks_carryForwardwriteRoll uncompleted scheduled and in-progress tasks from one day onto another - the evening-review step so nothing is silently missed. All arguments are optional: fromDate defaults to today and toDate defaults to tomorrow, both ISO dates like 2026-08-14.
tasks_clarifyTaskwriteTurn a raw inbox task into an actionable one; a task cannot leave the inbox until it has a next action. Requires task_id (format task_<epoch>, from tasks_getInbox) and nextAction (the single physical next step). Optional: definitionOfDone, title, priority (1 highest to 3 lowest), energy (low, medium, high, deep), contexts, deadline.
tasks_completeTaskwriteMark a task done and stamp completedAt on it. Use for 'I finished X' or 'mark X complete'. Requires task_id, the id from tasks_listTasks (format task_<epoch>, e.g. task_1786561481). If the task has a definitionOfDone, it is surfaced so you can check the work really is finished.
tasks_dropTaskwriteAbandon a task without completing it, keeping it on record as dropped rather than deleting it. Use for 'forget about X' or 'that's no longer needed'. Requires task_id (format task_<epoch>, from tasks_listTasks). Optional reason records why it was dropped.
tasks_getCandidatesreadReturn the tasks worth triaging now - due soon, gone stale, or clarified and ready to schedule. Use it to drive a morning planning routine. All arguments are optional: stale_threshold_days sets how many untouched days counts as stale (default 14).
tasks_getInboxreadReturn every task still sitting in the inbox, i.e. captured but not yet clarified. Use it to run an inbox review or answer 'what have I not sorted out yet'. Takes no arguments - call it bare.
tasks_getStaleTasksreadReturn tasks that have not been touched in a while (done and dropped ones excluded), so neglected work can be resurfaced or dropped. All arguments are optional: threshold_days sets how many days of no activity counts as stale (default 14).
tasks_listTasksreadList tasks, optionally narrowed by status, context or date - the general-purpose way to answer 'what's on my list'. All arguments are optional: status (inbox, clarified, scheduled, in_progress, done, dropped), contexts, dueBefore (ISO date), scheduledFor (exact ISO date).
tasks_rescheduleTaskwriteMove a task to a different date and put it in the scheduled state. Use for 'push X to Friday' or 'do X tomorrow instead'. Requires task_id (format task_<epoch>) and scheduledFor (an ISO date, e.g. 2026-08-14). The task must already have a nextAction set by tasks_clarifyTask.
tasks_startTaskwriteMark a task as in progress, to record that work has actually begun on it. Requires task_id, the id from tasks_listTasks or tasks_getInbox (format task_<epoch>, e.g. task_1786561481). Finish it later with tasks_completeTask.
tasks_updateTaskwritePartially update an existing task, changing only the fields you pass and leaving the rest untouched. Requires task_id (format task_<epoch>, from tasks_listTasks); every other argument is optional: title, nextAction, definitionOfDone, priority, energy, contexts, deadline, scheduledFor, parentId.
Telegram
Connects to a Telegram Bot via long-polling; maps Telegram user IDs to internal Kamari usernames for permission-gated event routing; provides tools to send messages and manage user links.
5 tool calls: 2 read, 3 write
telegram_get_linked_userreadLook up which Kamari username a Telegram account is linked to. Requires telegram_user_id: the numeric Telegram ID as a string. Use it to check whether an incoming Telegram user is known before acting on their message.
telegram_link_userwriteLink a Telegram account to a Kamari username, so messages from that Telegram user are treated as that user for permissions. Requires telegram_user_id (their numeric Telegram user ID as a string, 1 to 1099511627775) and username (an existing Kamari account belonging to a person; admin, plugin identities and service clients are refused). Optional display_name is a readable label. Replaces any existing mapping for that ID.
telegram_list_linked_usersreadList every Telegram account currently linked to a Kamari username, with its display name. Use it to find a telegram_user_id before unlinking, or to audit who can drive the bot. Takes no arguments - call it bare.
telegram_send_messagewriteSend a message to a Telegram chat or user via the bot. Requires chat_id (a numeric chat id or an @username) and text (the message body). Optional parse_mode enables formatting - MarkdownV2 or HTML - and is plain text when omitted.
telegram_unlink_userwriteRemove a Telegram-to-Kamari account link, after which messages from that Telegram user are no longer attributed to any Kamari user. Requires telegram_user_id: the numeric Telegram ID as a string, as shown by telegram_list_linked_users.
Terraform
Wraps the local Terraform CLI to plan/apply infrastructure from a sandboxed working directory, plus a one-call Proxmox VM/LXC provisioning convenience capability.
12 tool calls: 5 read, 7 write
terraform_applywriteRun terraform apply to create or modify real infrastructure. Safe by default: unless you pass approve=true it runs a plan-only dry run and changes nothing, so confirm with the user before approving. All arguments are optional: working_dir, var_file, extra_vars (key=value entries), approve.
terraform_destroywriteRun terraform destroy to tear down every resource this configuration manages. Safe by default: unless you pass approve=true it only shows what would be destroyed. This is irreversible once approved, so always confirm with the user first. All arguments are optional: working_dir, var_file, extra_vars, approve.
terraform_fmtwriteRun terraform fmt to rewrite the configuration files in a directory into Terraform's canonical style. It edits .tf files in place but never touches infrastructure. All arguments are optional: working_dir picks a subdirectory under config.terraform.working_dir, defaulting to that root.
terraform_initwriteRun terraform init to download providers and prepare a working directory - the first step before plan or apply in a directory that has not been initialised. All arguments are optional: working_dir is a subdirectory under config.terraform.working_dir, and empty (the default) uses that root itself.
terraform_outputreadRead the output values recorded in a Terraform state, e.g. an IP address or hostname produced by a previous apply. All arguments are optional: name returns one specific output (empty returns all of them) and working_dir picks the subdirectory under config.terraform.working_dir.
terraform_planreadRun terraform plan to preview what would change, without changing anything - always safe, and the right way to show the user what an apply would do. All arguments are optional: working_dir, var_file (a .tfvars path relative to it), extra_vars (key=value entries), target (a single resource address to limit the plan to).
terraform_proxmox_provisionwriteGenerate a Proxmox Terraform configuration for one VM or container and run init plus apply on it in a single call. Requires kind (vm or lxc), node (the Proxmox node name), template (a source VM template name for kind=vm, or a CT template volume id for kind=lxc) and name (the new machine's hostname). Optional: vmid, cores, memory_mb, disk_gb, network_bridge, working_dir, and approve - which must be true to actually provision, otherwise it only writes the config and plans.
terraform_start_vmwriteCreate a new Proxmox VM with almost everything defaulted - the quick path for 'spin me up a VM'. Requires only disk_gb (disk size in GB) and memory_mb (RAM in MB); node, template, name, cores and network bridge come from config.terraform defaults. Note this provisions for real immediately, with no dry-run step - use terraform_proxmox_provision when you need to plan first or control the details.
terraform_state_listreadRun terraform state list to show which resources Terraform currently tracks in its state - use it to see what actually exists before planning a change or a destroy. All arguments are optional: working_dir picks a subdirectory under config.terraform.working_dir.
terraform_validatereadRun terraform validate to check the configuration files for syntax and internal consistency errors, without contacting any provider or touching infrastructure. All arguments are optional: working_dir picks a subdirectory under config.terraform.working_dir, defaulting to that root.
terraform_workspace_listreadRun terraform workspace list to show the workspaces in a working directory and which one is selected. Workspaces keep separate state for the same configuration, e.g. dev and prod. All arguments are optional: working_dir picks a subdirectory under config.terraform.working_dir.
terraform_workspace_selectwriteSwitch to a Terraform workspace, creating it first if it does not exist, so subsequent plan and apply calls use that workspace's own state. Requires workspace: the workspace name, e.g. dev or prod. Optional working_dir picks a subdirectory under config.terraform.working_dir.
Text to Speech
Synthesizes text to audio using Kokoro (local/CPU), AWS Polly, or a silent stub
1 tool call: 0 read, 1 write
tts_synthesizewriteConvert text into spoken audio, returning a playable audio_url plus duration_ms. Use for 'say this out loud', 'read this to me', or to produce a voice clip. Requires text: the words to speak. Optional: provider (kokoro, polly, dud, or auto to follow config) and polly_resume_task_id to resume a timed-out Polly async job.
Threads
A reference index of communications (email + phone) — stores only the communication id, correlating them via the people store into unified per-contact/company/domain threads; full content is resolved from the owning source on demand
8 tool calls: 5 read, 3 write
threads_get_communicationreadFetch one communication in full - an email with its headers and body, or a call transcript. Requires comm_id: the communication id from any threads list tool (an email id, or <+E164>_<timestamp> for a call). Content is resolved from the owning plugin on demand, so check content_resolved in the result.
threads_list_by_companyreadList every recorded communication with anyone at a company, unioning all its contacts' addresses, domains and phone numbers, newest first. Use for 'what's our history with company X'. Requires company: the name to match against contacts' company field (case-insensitive, exact). Optional: limit. Read-only: call threads_sync_emails first if you need email that arrived since the last sync.
threads_list_by_contactreadList every recorded communication with one person - emails matched on their address and calls matched on their number - newest first. Use for 'what have I discussed with X'. Requires contact_id, the people-store person id from people_search (e.g. person1786561481). Optional: limit (default 50). Read-only: lists what is already indexed, so call threads_sync_emails first if you need email that arrived since the last sync.
threads_list_by_domainreadList every recorded email sent to or from an email domain, newest first - broader than threads_list_by_company because it catches addresses with no matching person in the people store. Requires domain, e.g. example.com (case-insensitive). Calls have no domain so this is email-only. Optional: limit. Read-only: call threads_sync_emails first if you need email that arrived since the last sync.
threads_list_recentreadA unified newest-first feed of recently recorded communications across every contact - use to answer 'what's come in lately'. All arguments are optional: limit caps how many come back (default 50) and kind filters to 'email' or 'phone' (empty means both).
threads_record_callwriteRecord a voice call transcript into the communications history so it appears alongside emails. Requires phone (the other party's number in E.164 format, e.g. +27821234567), timestamp (call time, ISO 8601 such as 2026-07-15T09:30:00Z) and transcript (the transcription text). Optional: direction, from_addr, to_addr, subject.
threads_record_emailwriteIndex one specific email into the threads history so it shows up in per-contact and per-company timelines. Requires email_id, the id returned by email_read or email_get. Only the id and correlation metadata are stored; the body stays in plugin.email. Use threads_sync_emails to bulk-index instead.
threads_sync_emailswriteScan the most recent emails from plugin.email and record any not yet indexed into the threads history - the bulk version of threads_record_email. All arguments are optional: limit sets how many recent emails to scan (default 500).
Time-Series Datastore
Local time-series store (metrics, timestamps, tags, range/aggregate queries, retention). Provider-swappable for a real TSDB backend.
7 tool calls: 3 read, 4 write
tsdb_aggregatereadDownsample a series into fixed-width time buckets - the way to chart or summarise data rather than list every point. Requires series (the series name from tsdb_list_series). Optional: fn (count, sum, min, max or avg; default avg), bucket_seconds (default 60), from_ts, to_ts, tags, provider.
tsdb_delete_pointswritePermanently delete data points from a series within a time range and optional tag filter -- more targeted than tsdb_prune. Requires series, from_ts and to_ts (inclusive ISO or epoch bounds, seconds or milliseconds). Optional tags narrows the delete to points matching all given key=value pairs; omit to delete every point in range. This cannot be undone.
tsdb_list_seriesreadList every time-series name the store knows about, so you can discover the valid series value for tsdb_query, tsdb_aggregate or tsdb_prune. All arguments are optional: provider overrides the configured backend for this call.
tsdb_migratewriteCopy all time-series data from one storage backend to another, e.g. when moving off the local store. Requires from_provider (the backend to read from) and to_provider (the backend to write into). Both must be configured provider names.
tsdb_prunewritePermanently delete old data points from a series to reclaim space. Requires series (the series name from tsdb_list_series) and before_ts (an ISO datetime or epoch; every point strictly before it is deleted). This cannot be undone.
tsdb_queryreadFetch the raw recorded data points for one series, optionally narrowed by time range and tags. Use for 'show me the readings for X'. Requires series: the series name, as listed by tsdb_list_series. Optional: from_ts and to_ts (inclusive bounds, ISO or epoch), tags, limit (default 1000), provider.
tsdb_writewriteRecord one or more time-series data points, e.g. a temperature reading or a counter sample. Requires points: an array where each entry is {series, ts (ISO datetime or epoch), value (a number), tags (optional object)}. Optional provider overrides the configured backend for this call.
TLS Certificate Renewal
Obtain and renew Let's Encrypt TLS certificates via DNS-01 challenge on AWS Route53
2 tool calls: 0 read, 2 write
cert_issuewriteObtain or renew a TLS certificate for another service's own fully-qualified domain name (DNS-01 challenge on Route53), reusing this plugin's configured email, hosted zone, and AWS credentials. Returns the fullchain certificate and private key as base64 (cert_b64/key_b64) so the caller can install them without filesystem access to this plugin. Use for 'get me a cert for <fqdn>' from another service in the same DNS zone. Required: fqdn. Optional: force=true to renew even when the current certificate is still valid; ip=<IPv4 or IPv6 address> to also point the fqdn's A/AAAA record at that address in the same call (ip_updated in the result confirms it).
cert_renewwriteCheck the configured domain's TLS certificate and renew it from Let's Encrypt (DNS-01 challenge on Route53) if it is close to expiring. Use for 'renew the certificate' or 'is my cert about to expire'. All arguments are optional: pass force=true to renew even when the current certificate is still valid.
UI and Fallback Config
Handles config updates from web UI and speech/text fallback
2 tool calls: 1 read, 1 write
config_getreadRead a plugin's currently stored configuration plus its version number. Use it before config_update so you can send back the full object and the matching expected_version. Requires path: the registry path to read, e.g. config.tts or config.vision. config.vault and config.vault.key (the vault's own storage) are always refused; use vault_get. config.auth.*, config.oauth and config.oauth.* (sessions, API keys, users, OAuth settings, signing keys, clients and tokens) are refused too; use the auth_admin_* tools, or GET /auth/oauth/config for the OAuth settings.
config_updatewriteWrite a plugin's configuration into Kamari's config registry, replacing that path's whole config object. Requires path (a registry path starting with 'config.', e.g. config.tts) and value (the complete config object - any key you leave out is dropped, so read it with config_get first). Optional expected_version: pass the version from that config_get so the write fails instead of clobbering a concurrent change. A value that breaks the owning plugin's config_schema (wrong type, not in enum, outside minimum/maximum) is refused with an error naming the field and reason. config.vault and config.vault.key (the vault's own storage) are always refused; use vault_set. config.auth.*, config.oauth and config.oauth.* (sessions, API keys, users, OAuth settings, signing keys, clients and tokens) are refused too; use the auth_admin_* tools, or PUT /auth/oauth/config for the OAuth settings.
Vision
Detects faces in an image and re-identifies people across photos. Unknown faces can be assigned a name; once named, the same person is recognised automatically in future images. Also detects exposed private body regions (NSFW) with bounding boxes, and supports teach-by-example custom object detection (label a region, then match new images against it). Also decodes QR, Data Matrix, UPC-A/E, EAN-13/8, Code 128/39/93, PDF417, ITF, and Aztec barcodes, and reads printed English text (OCR) with bounding boxes. Runs small ONNX models (SCRFD + ArcFace + NudeNet + DINOv2 + YOLOv8n + PP-OCRv3) on CPU — no cloud, no GPU.
17 tool calls: 6 read, 11 write
vision_analyzereadAnalyse an image: identify faces against the known-people gallery, flag exposed body regions, match custom taught objects, decode barcodes and QR codes, and read printed text. All arguments are technically optional, but you must supply the image as either image_path (a file under the data dir) or image_b64 (base64 image bytes) or there is nothing to analyse. Optional detections picks which to run as a list or comma-separated string - any combination of face, nsfw, object, barcode, text - and defaults to face only. Optional: max_faces, min_score. Records nothing, so faces come back with a null face_id; to enrol a face, use vision_analyze_and_record instead.
vision_analyze_and_recordwriteSame analysis as vision_analyze, but also stores every detected face on disk as a sighting (its face embedding, box and score, plus a thumbnail) so it can be enrolled. Use it only when you intend to call vision_enroll_face on the result: it writes biometric data each time it finds a face. Arguments are identical to vision_analyze - supply the image as image_path or image_b64; optional detections, max_faces, min_score. Each face's face_id is the value vision_enroll_face needs.
vision_enroll_facewriteTeach the system who a face belongs to, so that person is recognised automatically in future images. Requires face_id, taken from a vision_analyze_and_record result (vision_analyze records nothing and returns no face_id). Also pass either name (matches an existing person case-insensitively, or creates a new one) or entity_id from vision_list_people (adds to that exact person, typo-proof, and wins if both are given).
vision_forget_objectwritePermanently delete a custom object class along with every labelled example and thumbnail, after which it is never detected again. Requires object_id: the class id from vision_list_objects. This cannot be undone - use vision_set_object_enabled with enabled=false to mute a class instead.
vision_forget_personwritePermanently delete an enrolled person and every face sample belonging to them, after which they are no longer recognised in any image. Requires entity_id: the person's id from vision_list_people. To remove just one bad face instead, use vision_unenroll_face.
vision_gallery_versionreadReturn the current face+object gallery version counter - a cheap poll that loads no images or galleries. Compare it against the version tagged on a photo's earlier vision_analyze result to decide whether re-analysing that photo could now produce a new match. Takes no arguments - call it bare.
vision_get_object_exemplarsreadShow every labelled example image for one custom object class as base64 JPEGs, with its current thresholds and the similarity spread across examples - use it to spot a mislabelled example when a class matches badly. All arguments are optional, but pass either object_id (from vision_list_objects) or label (case-insensitive exact match) or it cannot identify a class.
vision_get_person_facesreadShow every enrolled face thumbnail for one person as base64 JPEGs, each with its face_id, for checking whether a wrong face got attached to them. All arguments are optional, but pass either entity_id (from vision_list_people) or name (case-insensitive exact match) or it cannot identify a person.
vision_label_objectwriteTeach a custom object class by example: mark a region of an image and give it a name, and vision_analyze can then spot that object in new images. No model training happens. Requires box (the region as [x1, y1, x2, y2] in pixels, or a comma-separated string like '10,20,110,140') and label (the class name, e.g. tent). Also supply the image via image_path or image_b64.
vision_list_objectsreadList the custom object classes taught so far, each with its object_id, how many labelled examples it has and whether it is currently enabled for detection. Use it to find the object_id the other object tools need. All arguments are optional: search filters labels by case-insensitive substring.
vision_list_peoplereadList everyone enrolled in the face gallery, with how many face samples each has and their entity_id. Use it to find the entity_id needed by the other people tools, or to answer 'who do you recognise'. All arguments are optional: search filters names by case-insensitive substring.
vision_merge_objectswriteFold one custom object class into another, moving all its examples across and then deleting it - for when the same real object was taught as several separate classes. Requires source_object_id (the class to merge from, which is deleted) and target_object_id (the class to merge into, which survives and keeps its label). Both ids come from vision_list_objects.
vision_rename_objectwriteChange a custom object class's label, leaving its examples and matching untouched - future detections simply report the new label. Requires object_id (the class id from vision_list_objects) and label (the new class name). A label that collides with a different existing class is rejected; use vision_merge_objects to combine two classes instead.
vision_rename_personwriteChange an enrolled person's display name, leaving their face samples and recognition untouched - future matches simply attach to the new name. Requires entity_id (the person's id from vision_list_people) and name (the new display name). A name that collides with a different existing person is rejected.
vision_set_object_enabledwriteMute or un-mute a custom object class for detection while keeping all its examples - the right tool for a class that fires too often, since vision_forget_object destroys its examples permanently. Requires object_id, from vision_list_objects. Optional enabled defaults to true; pass false to mute. Adding a new example does not re-enable a muted class.
vision_unenroll_facewriteRemove one face sample from a person - the fix when a face was matched to the wrong individual. Requires face_id (the face or sighting id to remove). Optional entity_id limits removal to that one person; empty removes it from everyone holding it. To move a face, unenroll it then call vision_enroll_face with the right person; a person left with no faces is deleted.
vision_unlabel_regionwriteRemove one labelled example from a custom object class - the fix for a region that was labelled wrongly and is hurting matching. Requires exemplar_id, as shown by vision_get_object_exemplars. Optional object_id limits removal to that one class; empty removes it from every class holding it. A class left with no examples is deleted.
Webhook
Inbound webhook endpoints — each named webhook receives HTTP requests and emits them as events pipelines can trigger on
5 tool calls: 2 read, 3 write
webhook_createwriteCreate an inbound webhook endpoint and return its secret URL, so an outside system can push events into Kamari for pipelines to trigger on. Requires name: a unique internal handle of letters, digits, _ and - . Optional: label describes what calls it, methods lists accepted HTTP verbs (default ["POST"]). To require signature verification, also pass signing_secret plus signature_header and timestamp_header naming the exact request headers your sender uses (no default header names are assumed). Treat the returned URL token, and any signing_secret, like a password.
webhook_deletewritePermanently delete a webhook; its URL starts returning 404 immediately and any system still calling it breaks. Requires webhook_id: the id from webhook_list. The secret token cannot be recovered afterwards, so recreating it means re-issuing a new URL.
webhook_getreadFetch one webhook including its secret token, signing secret, and URL path, for re-copying credentials into a calling system. All arguments are optional, but pass either webhook_id (from webhook_list) or name (the internal handle, matched case-insensitively) or it cannot identify a webhook.
webhook_listreadList every inbound webhook with its delivery stats and whether signature verification is on, to find a webhook_id or check whether one is actually receiving calls. Secret tokens and signing secrets are withheld here - use webhook_get to reveal them. All arguments are optional: search filters over name and label, case-insensitively.
webhook_updatewriteRename a webhook, change its label or accepted HTTP methods, or turn signature verification on/off, leaving its URL and secret token untouched so existing callers keep working. Requires webhook_id, from webhook_list. Optional and omit-to-keep: name, label (pass "-" to clear it), methods (replaces the whole set), signing_secret/signature_header/timestamp_header (pass "-" on signing_secret to disable verification), max_signature_age_s.