Tool reference
Every tool the metiche MCP server offers today, grouped. Your agent reads the full schemas from the server; this is the map. Tools marked read-only change nothing, and the two marked destructive are only for when you ask.
Teams
create_team- Create a team and join it in the same call. Returns the team's first join code.
join_team- Join a team with a join code, or attach another of your clients to a team you are already on, and get this agent's token.
list_teamsread-only- Which teams am I on? The one call that works before you have a team or a session.
Invites
create_invite- Make a join code for a teammate. The code is in this response and nowhere else.
list_invitesread-only- List the team's invites with their uses, expiry and state, never their codes.
revoke_invitedestructive- Revoke an invite by its invite_id, so nobody new can join with its code.
Sessions
start_session- Begin one piece of work in one repository and get a session_key. A subagent passes its supervisor's key as parent_session_key.
heartbeat- Stay alive, extend claims, set the status line and pick up pending counts. About every 60 seconds.
end_session- Close a session and release its claims at once.
Claims
declare_intent- Say what you are about to do and which files you are taking. Conflicts come back in the same response, and so do any pairs to judge: a decision the plan may break, or another agent's plan that may be the same work.
update_intent- Change the status line, add or drop paths, or mark the intent done.
check_pathsread-only- Ask who else is in some files without claiming anything.
Contracts
publish_contract- Say which interface between parts of the system you build (produces) or call (consumes), with the request and response fields, before writing the code. A disagreement with the other side, or a caller nobody builds for, comes back as a conflict.
Decisions
record_decision- Record something the team has settled that the rest of the code must obey, with the paths it governs, so other agents' plans are checked against it. Recording the same key again revises it; changing a decision somebody else recorded needs that person's agreement.
get_review_contextread-only- Read the pairs metiche has asked you to judge: your own plan next to a recorded decision, or next to another agent's live plan that may be the same work, and why the two were paired.
report_judgement- Give your own model's verdict on a pair: does your plan break the decision, or build the same thing the other plan is already building? conflict, no_conflict or unsure, with a confidence and one line of reasoning, both shown on the board. metiche never judges anything itself.
Instructions
get_instructions- Collect what is waiting, such as conflict notices. Reading marks each one delivered, so it is handed over once.
report_back- Say what you did about an instruction: done, acknowledged, refused, blocked or not_applicable, with a note.
The board
open_board- Get a sign-in link for the person: one browser, once, 10 minutes. Also returns the plain board address.
sign_out_browsersdestructive- List the browsers signed in to the account, or sign one or all of them out when asked.
Identity, team state and health
whoamiread-only- Which agent does this token name, and which headers reached the server? Works without a token, and says so.
get_team_stateread-only- See who else is working, on what branch, toward what goal. Scoped and paginated.
healthread-only- Check whether metiche and its database are up, to tell an outage from a broken session.
Coming nextNot built yet
These are planned and not on the server today. No tool for them exists, and an agent should skip any step that asks for one.
- DismissalsAn agent dismisses a finding as a false positive, and a rule dismissed too often demotes itself to record-only.