TuringCorp — MCP server

A decision model for the calls that don't have a right answer. Your agent has two defensible options and has to pick one. Send both: a panel of models judges them together and returns the better one, how far apart it judged them (a calibrated confidence), and why.

• Endpoint (canonical): https://mcp.turingcorp.net/mcp
• Aliases: https://mcp.turingcorp.net/ and https://mcp.turingcorp.net/mcp/
• Health: https://mcp.turingcorp.net/healthz
• Transport: Streamable HTTP, stateless (no session, no Durable Object). GET /mcp → 405 is normal.
• Protocol: MCP 2026-07-28 (modern) and 2025-06-18 (legacy) — one route serves both.
• Discovery methods (server/discover, tools/list) need no credentials — directory scanners can read the full tool list unauthenticated, by design.

Making a decision: decide

A successful call returns the decision inline. job_id, betterOption, confidence and reason all arrive in the same tool result — there is nothing to poll and nothing to fetch afterwards. The job_id is only for the case where the call never came back (timeout, dropped connection, client gave up waiting).

confidence is this service's own judgement of how far apart the two options were — a reference for your decision-making, not an instruction, not a result, and not a prediction of how the choice turns out. As a rough reading: ≥90% clearly apart · 80–90% apart, less clearly · 70–80% a closer call · <70% close to evenly matched. These ranges are descriptive only. Choose your own threshold for your own use case; for high-stakes or irreversible decisions apply your own review policy. Observed accuracy for each range, and how it was measured, is published at https://api.turingcorp.net — that page is the authority on the bands and their measured accuracy.

⚠️ idempotentHint: false is an honest declaration: a retried call is a new call. Record the job_id the result returns — a decision you have already paid for can be collected for 7 days with the same credential (GET https://api.turingcorp.net/v1/jobs?job_id=<id>; without the id, GET https://api.turingcorp.net/v1/jobs lists the job ids for that credential). Retrieve instead of retrying.

When not to use it

• More than two options. It compares exactly A and B — there is no third slot, and it will not rank a list.
• Anything you can compute or verify. A spec, a test, a price, a document: if something objective decides it, use that. This is for choices where no objective rule does.
• Factual questions. It picks between two candidates; it does not look anything up.
• Speed-critical paths. A decision is a long call and a paid one. Do not put it behind a request that has to answer in seconds.
• High-stakes irreversible calls without review. Route on the confidence and keep your own review policy.

How to fill the three arguments

• task — state the decision neutrally, without leaning toward either side: "Which email do I send?", not "Should I send the honest one?"
• optionA / optionB — one concrete option each, plus the case for it. Plain text or Markdown, any length; keep the two sides roughly comparable so the comparison is fair.
• One option = one plan. Do not bundle alternatives into a single side ("go indoors or postpone"): it compares the two slots, it does not split one of them for you.

{
  "task": "Which version of the delivery-slip email do I send to a client we want to keep?",
  "optionA": "Short and direct: the integration took longer than planned, delivery moves to the 24th, everything else is unchanged.",
  "optionB": "Warmer and longer: thank them for the kickoff, explain that dependencies took more time, offer to walk through the details."
}

More patterns — choosing the version you send, thresholding on the confidence, and writing the two options so the comparison is fair: docs/COOKBOOK.md.

Retrieving a result

Every call returns a job_id. A call that is allowed to finish returns the decision inline — the job_id is only for the case where it never came back.

⚠️ How to retrieve — use get_result. Retrieval needs the Agent Pass, and inside an MCP host the agent usually does not have it — so retrieval is a tool call, not a REST call: get_result is authenticated by the host, which attaches the pass for you. An agent never handles the pass, and the credential is deliberately not a tool argument (it would end up in prompts and transcripts).

• get_result with a job_id — that job's status, and once it succeeded, the same decision body the original call returned.
• get_result with no argument — the job ids this credential created in the last 7 days. This is the case that matters most: when a call is cut off, you never received an id.
• Read-only and free — it starts no new work and costs nothing, so it is safe to retry (readOnlyHint: true, idempotentHint: true).
• A host that declares the io.modelcontextprotocol/tasks extension can poll tasks/get instead. Most clients do not declare it yet, which is why get_result exists.

Operator/host, with the pass (the same data over REST):

• With the id — GET https://api.turingcorp.net/v1/jobs?job_id=<id> using the same Agent Pass.
• Without it — GET https://api.turingcorp.net/v1/jobs lists the job ids that credential created in the last 7 days; then fetch one as above. The list carries only the job id, the product and a timestamp.
• Nothing yet — an empty list is not an error, and there is no job_id=0 placeholder: {"object":"list","window_seconds":604800,"data":[]}. Asking for 0 returns 404 No such job.

Retrieval returns the job's status and, once it succeeded, the same body the call itself would have returned. A job that is not yours, or older than 7 days, is reported as unavailable. GET https://api.turingcorp.net/v1/account returns {"account_id":"…"}.

The practical answer is not to need retrieval: reserve 180–300 seconds so the call is allowed to finish.

See it decide first

There is no free tier, so we publish the evidence instead: 27 real decisions, recorded verbatim — the question, both options, which one was preferred, the confidence reported, and the reason. Nine domains, three cases each: Writing · Tech · Business · Research · Career · Money · People · Travel · Everyday.

→ https://github.com/TuringCorp-net/poe-demo-public — dataset: examples.json

Read the confidence column first. On ordinary, closely matched questions it reports 70–90%, not 99% — and that is the point of a calibrated number rather than a decorative one: a high value means the comparison was decisive and the pick can be acted on, a low value means the two options really are close and the choice stays yours. In this set the range is 27.3%–88.3%, median 74.0%, none above 90%, because these are everyday close calls, not easy ones. What the bands deliver when they are decisive is published at https://api.turingcorp.net.

Authentication

Send an Agent Pass: Authorization: Bearer <pass>.

Issued at https://agent-pass.turingcorp.net — self-service signup (email verification) → top up → get a pass. A pass is valid for 7 days and can be re-rolled. If a call is refused as an invalid credential, sign in there again and re-roll.

Errors fall into two classes, and they arrive in different places:

Credential problems are rejected before the call (HTTP 401 + WWW-Authenticate). Business failures come back as a tool result with isError: true plus a JSON block {error, http_status, action_url, message} — http_status is the upstream status; the tool call itself is HTTP 200.

Branch on the HTTP status first: a 401 will not succeed on retry without a new credential, whereas a business failure was a call that ran and was declined — read the block for the reason and the next step. The detailed contract is under Errors are machine-readable below.

Entry rate limit: 120 requests / 60 s / client IP. This is flood damping, not a quota — the real per-key quota is enforced separately. The counter is approximate: it is maintained per edge location and is eventually consistent, so a brief overshoot is possible.

Two ways in — and they are not the same