API and MCP

The studio's judge, without the AI, at https://api.ztljudge.com — no key, no sign-up. In a row, status is verified, refuted or unverified (T, F and Z in the studio).

Limits and records: 30 requests a minute per address. Each request is logged for 30 days with the address, the client's User-Agent, the endpoint, the status and a SHA-256 fingerprint and size of what was sent — not the document itself and not the answer. The server runs the open-source code of the studio; GET /health reports the commit, and you can run the same judge yourself (see Code and proofs).

HTTP

curl -s -X POST https://api.ztljudge.com/judge \
  -H 'Content-Type: application/json' \
  -d '{"claim": "p -> p", "rows": [{"name": "p", "means": "it rains", "status": "unverified"}]}'

The answer carries the disposition, what it means, and which names to check. GET /language describes how to write a table, GET /examples returns worked documents, GET /health reports the commit of the code that answers.

For AI agents (MCP)

https://api.ztljudge.com/mcp (Streamable HTTP) offers three tools: language, examples and judge. Your agent's own model writes the table; ZTL judges that table. Whether the table says what was meant is not checked by the judge — with no person looking, nobody checks the table, so treat the verdict as a verdict on what your agent wrote. As with any remote tool, treat its answers as data, not instructions. To add it to Claude Code:

claude mcp add --transport http ztl https://api.ztljudge.com/mcp

The server is listed in the MCP registry as io.github.inventor1975/ztl-judge; a listing there is not a review.