Check’d in your AI assistant

The Check’d MCP server lets an AI coding assistant work with your test results: look at the latest run, read a failure’s exact request, response and diagnosis, fix the code, and re-run the failed tests to confirm. It runs on your machine over stdio and talks to the Check’d API with your key.

Tools that only read are free. Tools that start runs, scans or re-runs use your workspace’s allowance, and say so in their descriptions so the assistant asks before using them.

1. Get a key

Create a workspace API key under Settings, API keys (owners and admins). The key reaches only that workspace. If you already ran checked login on this machine, the server can use that sign-in instead and you can leave the key out.

2. Add it to your assistant

Claude Code

claude mcp add checked --env CHECKED_API_KEY=tfy_your_key -- npx -y @checked/mcp

Cursor: .cursor/mcp.json in your project, or ~/.cursor/mcp.json

{
  "mcpServers": {
    "checked": {
      "command": "npx",
      "args": [
        "-y",
        "@checked/mcp"
      ],
      "env": {
        "CHECKED_API_KEY": "tfy_your_key"
      }
    }
  }
}

VS Code: .vscode/mcp.json (VS Code asks for the key once and stores it)

{
  "inputs": [
    {
      "type": "promptString",
      "id": "checked-api-key",
      "description": "Check’d workspace API key",
      "password": true
    }
  ],
  "servers": {
    "checked": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@checked/mcp"
      ],
      "env": {
        "CHECKED_API_KEY": "${input:checked-api-key}"
      }
    }
  }
}

Needs Node.js 20 or newer; npx fetches the server the first time. Set CHECKED_API_URL as well if you use a Check’d server other than https://getchecked.twelveai.app.

Tools

Tool	What it does
whoami	Read-only, free. Which account or key the server uses.
list_projects	Read-only, free. Projects with their last run.
list_environments	Read-only, free. A project’s environments (secrets masked).
list_runs	Read-only, free. Recent runs.
get_run	Read-only, free. Status, release gates, verdict, top problems.
wait_for_run	Read-only, free. Wait up to 10 minutes for a run to finish.
get_run_results	Read-only, free. Individual API test results.
get_run_report	Read-only, free. The AI QA report.
list_findings / get_finding	Read-only, free. Problems with request, response, stack trace, diagnosis and fix.
list_issues / get_issue	Read-only, free. Issues (today the latest run’s findings).
get_usage	Read-only, free. Session and month usage.
preview_import	Read-only, free. What an API import would add (nothing saved).
start_run	Starts work, uses allowance. Start a run, optionally wait for the result.
rerun_failed	Starts work, uses allowance. Re-run only the failed tests of a run.
start_scan	Starts work, uses allowance. Scan a website.
import_spec	Makes changes. Save endpoints from an API document.
cancel_run	Makes changes. Cancel a run.

Projects and runs are named by their ref or number; finding ids look like r123 (a failed API test) or f456 (a website finding). start_run and start_scan can wait for the result for up to 10 minutes; for longer runs the assistant calls wait_for_run again rather than starting another run.

Resources and prompts

Run reports, run results and findings are also readable as resources: checked://runs/{run}/report, checked://runs/{run}/summary and checked://findings/{id}. The fix_failures prompt walks the assistant through the latest run of a project and fixes what it can in your code.

Things to ask

What failed in the latest Check’d run of payments-kqzxm, and why?
Show me the request and response for the worst finding and fix the handler.
I fixed it; re-run the failed tests and tell me if the release gates pass now.
How much of this session’s Check’d allowance is left?

Usage and safety

When the allowance for the session is used up, tools that start work return the server’s message and the reset time, and the server tells the assistant not to retry. Environment settings come back with their secrets masked, as in the app. Read-only tools do not start work or change your projects. Test evidence (requests and responses) is shown as recorded, as it is on the run page, so treat it like any test log.