Aller au contenu principal

Connect an AI agent to CheckFox over MCP and let it work through an audit.

Cette documentation est disponible en anglais uniquement.

Goal

Let an MCP client read an audit, follow the official methodology criterion by criterion, and write verdicts you review.

Preconditions

  • The audit's own workspace has MCP (included from Pro, add-on on Free and Starter). The check is per audit.
  • The audit has its samples. The agent evaluates criteria, it does not sample (methodology / sampling).
  • The agent acts as you: it cannot reach an audit you cannot, nor write where you only read.

Connecting

With the native connector, no key

  1. In your MCP client, add a custom connector.
  2. Name: CheckFox. URL: https://api.checkfox.eu/mcp. OAuth client id and secret: empty.
  3. Add it. CheckFox opens sign-in and consent.
  4. Sign in, choose read-only or also record verdicts, approve.

With an API key

  1. In User Settings > Integrations, generate an API key. Tick mcp:read to read audits and methodologies; mcp:write only to record verdicts.
  2. Copy the key: shown once, it is the security boundary, treat it as a password. Generating asks your password if not entered in the last 5 minutes. Keys expire one year after creation; generate a new one before.
  3. Configure the client to call https://api.checkfox.eu/mcp with the key as bearer token. Snippets for common clients: checkfox.eu/mcp.

How the agent then works

One sample, one criterion at a time:

  1. get_next_criterion returns the next untested criterion with its official tests; it can claim it so parallel agents do not collide.
  2. The agent picks a method: read the source, drive a browser, or run the scanner for an extra signal.
  3. It gathers evidence against every test.
  4. It decides a status, with a user-impact level on failure.
  5. update_criterion writes status, findings and suggested fix; next criterion.

Verify success

  • Audit progress advances; new entries are marked machine-written.
  • Criteria you decided are unchanged: writes only land on "Not tested".
  • Where the agent could not conclude: Needs review, not Success.

Review the result

An agent pass is a first pass, not a delivered audit. Before the report or statement: check ambiguous criteria, rewrite problem and solution text in your words, add evidence images it cannot capture. See methodology / finding quality.

Common pitfalls

  • Granting mcp:write when read would do. Give the key only what the agent needs.
  • Expecting it to sample for you. Add samples and decide scope yourself.
  • Running it on a workspace without MCP. Tools refuse per audit with an explicit message, even with a valid key.
  • Publishing straight from an agent pass. Every verdict needs a human reading.

Dernière révision : 2026-11-02