김인턴
Tools

Tool catalog

Every capability tool, by namespace, generated from the canonical catalog.

The tables below are rendered at build time from the catalog the agent itself reads (pkg/capabilityprotocol/generated/capability-tools.json, protocol 0.4.0), so they cannot drift from what is deployed. The descriptions are the exact English sentences the model receives.

Tools whose effect column is destructive, plus message_send and channel_update, pause for approval before running.

Tasks

ToolEffectWhat the model reads
task_addworkspace_writeCreate a new workspace task with typed task fields.
task_listreadList workspace tasks with optional filters.
task_updateworkspace_writeUpdate explicit fields on an existing task, including who takes part in it.
task_deletedestructivePermanently delete a task.

Backed by admind's Flow API (internal/capabilityd/flow_tool.go). Updates and deletes take a taskHint: the exact task ID or exact current title from a task_list result. Dates are yyyy-mm-dd; the agent resolves "next Friday" before calling.

People

ToolEffectWhat the model reads
person_listreadList the people in this workspace with their exact name, email, @handle, and mention.

Implemented in internal/capabilityd/person_list_tool.go over the same Flow membership data. The agent calls it whenever a message names someone partially, so every later hint carries an exact name, @handle, or email.

Calendar

ToolEffectWhat the model reads
event_addworkspace_writeCreate a calendar event with a concrete time range.
event_listreadList calendar events in a concrete time window, optionally filtered by title, description, or location.
event_updateworkspace_writeUpdate explicit fields on a calendar event.
event_deletedestructivePermanently delete a calendar event.

Implemented in internal/capabilityd/calendar_tool.go, also over admind's Flow API; a calendar event is a task with a time span. Times are ISO timestamps with a time zone, and eventHint follows the same resolution rule as taskHint.

Messages

ToolEffectWhat the model reads
message_contextreadReturn the exact current Mattermost conversation, thread, requester, and bot identities.
message_searchreadFind messages in an exact conversation scope, or read known ones in full.
message_sendexternal_sendSend a Mattermost message to a direct message, channel, or the current conversation after approval.
message_updateexternal_writeReplace one exact span of text inside your own earlier Mattermost message, leaving the rest of it untouched.
message_deletedestructivePermanently delete exact Mattermost message IDs from message_search after approval.

Implemented in internal/capabilityd/platform_message_tool.go and mattermost_tool.go. message_search has two modes: keyword queries return IDs with a preview around the match, and messageIDs returns those messages complete. message_update edits one exact span of the agent's own earlier message: oldText must occur exactly once, and the rest of the message stays untouched. message_send targets a direct message, a named channel, or the current conversation; only the current-conversation targets skip approval.

Channels

ToolEffectWhat the model reads
channel_updateexternal_writeUpdate an exact Mattermost channel display name, header, or membership after approval.

Implemented in internal/capabilityd/mattermost_tool.go. Display name, header, and membership changes all go through this one gated tool.

Web

ToolEffectWhat the model reads
web_searchreadSearch the public web and return ranked result snippets.

Implemented in internal/capabilityd/web_tool.go through OpenRouter's web search. Domain allowlists, language, and location narrow the query.

Sites

ToolEffectWhat the model reads
site_servesite_publishServe a site project directory you built in the workspace: preview mode returns a temporary review URL, publish mode deploys to the public URL.
site_listreadList served sites with their exact siteID, slug, lifecycle status, and published URL.
site_unservedestructiveTake a served site down after explicit runtime approval: unpublishes it, frees its slug, and deletes the server-side record.

Implemented in internal/capabilityd/site_tool.go over admind. The agent builds a site as ordinary files in the workspace, then serves that directory: preview returns a temporary review URL, publish takes the public one. The slug is derived from the title on first serve and owned by the server. site_unserve frees the slug and deletes the server-side record; the source files in the workspace stay.

Documents and images

ToolEffectWhat the model reads
document_readreadRead a workspace document from an exact /workspace path and return Markdown content.
ToolEffectWhat the model reads
image_readreadRead a workspace image from an exact /workspace path and return a base64 attachment.

Implemented in internal/capabilityd/file_read_tool.go (with file_read_helper.py doing format extraction). These read a workspace path and return Markdown or a base64 attachment, which is how a PDF someone uploaded becomes something the model can quote.

Browser

ToolEffectWhat the model reads
browser_openconnectOpen an exact HTTP or HTTPS URL in the available browser and return the resulting page identity and initial structure.
browser_snapshotreadRead the current browser page structure and return stable interactive references for inspection and control.
browser_screenshotreadCapture the visible browser page and upload it to a temporary workspace-visible device path for visual review.
browser_clickexternal_writeClick one exact target from the current browser snapshot and return the completed action.

Implemented in internal/capabilityd/browser_tool.go. Calls route to the companion on the user's own computer when it is connected, with the device browser as the other provider. browser_open requires the user present, and responses carry no cookies, profile paths, or local screenshot paths; browser_screenshot uploads to a temporary workspace-visible path instead. Login, MFA, and captcha steps are handed to the human.

Artifacts

ToolEffectWhat the model reads
artifact_reviewreadReview rendered artifact screenshots against a concrete intent and rubric, returning typed visual issues and suggested fixes.

Implemented in internal/capabilityd/artifact_review_tool.go. This is the agent's own quality loop: it renders what it built, screenshots it, and asks for typed visual issues against the stated intent before delivering.

On this page