The Claros MCP server lets an agent read the rules before it starts and call check_action before it saves. It’s the way in that works today without early access. Unlike the guard, the agent decides whether to ask.

​
Connect

​
Claude.ai, ChatGPT and Claude Desktop connectors

Not live yet: sign-in for connectors needs a one-time switch on the Claros side. Until then, use a key as above.
Connectors sign you in with your Claros account instead of a key (OAuth 2.1, as MCP specifies).
  • Claude.ai / Claude Desktop: Settings → Connectors → Add custom connector. URL: https://claros.sh/mcp. Leave the OAuth client fields empty.
  • ChatGPT (developer mode): add a custom connector with the same URL and authentication OAuth.
Then sign in on the Claros sign-in page and allow access. The connector reads and checks in your first workspace, with the rights of an agent key. No workspace yet: create or join one in the web app first, or the server answers 403. Removing you from the workspace cuts the connector off on its next call. Early-access users get the same tools through npx getclaros mcp, a stdio bridge init adds as the claros entry, using the saved key.

​
Tools

result is allowed, nudge, ask or stop, and guidance says in one line what to do. allowed: false (stop or ask): don’t proceed; each violation has the rule, the quote, who_to_ask and next (for ask: check with that person first). nudge: go ahead; each nudge carries the rule, the quote and, when there is one, a fix to apply. not_evaluable: facts the check needed and didn’t get; send them and check again. judge_yourself: fuzzy rules the server judge didn’t run; read them and ask if unsure. 401: key missing or wrong (or the connector’s sign-in expired: reconnect). 403: not an agent key, or a connector user in no workspace.

​
Skill export

GET /api/export/{workflow_id}.skill.md returns the map as a SKILL.md: steps in the owner’s words, rules as stop conditions. Unconfirmed parts are flagged.