Question 1

Is the HTTP transport safe to put on a network?

Accepted Answer

Not on its own. The README carries an explicit security warning: the endpoint is unauthenticated, and production remote deployments must be fronted by an authenticating OAuth 2.1 proxy per the MCP 2026-07-28 authorization specification. Out of the box it speaks stdio, which needs no such treatment. If you do enable HTTP, JUDGE_TRANSPORT must be exactly stdio or http and anything else exits with an error, HTTP_HOST defaults to loopback, and HTTP_PORT defaults to 8080 with a 1-65535 range check.

Question 2

I have several TypeSafe keys. What happens when one hits a rate limit?

Accepted Answer

It rotates before you get a fallback envelope. A comma-separated TYPESAFE_API_KEYS is tried in order with rotation on 429 and 529, and it takes precedence over TYPESAFE_API_KEY and the numbered TYPESAFE_API_KEY_2 through _9 variables, which fill in after the single-key variable when TYPESAFE_API_KEYS is unset. This matters because a fallback envelope is a silent degradation: the judge call returns an error payload instead of an answer, and your workflow continues with no decision made.

Question 3

Do I have to pick my own threshold on the noul answers?

Accepted Answer

Yes. The README is direct about it: a noul answer is a probability, not a verdict, and you should pick your own action threshold in the caller, for example proceeding only above 0.7. The response example shows ready_to_merge at 0.93 as a number the calling code has to interpret. Choice answers carry a confidence plus the full probability distribution, which is what makes routing and review queues possible without parsing prose.

Question 4

Why an MCP server rather than calling the API from inside my agent?

Accepted Answer

Portability and surface area. The same config shape works for Claude Desktop, codex, goose and pi, and there is one command for Claude Code via claude mcp add. The server contributes key discovery and rotation, the fallback envelope contract so judge failures do not break a workflow, cancellation of in-flight requests when the client disconnects, and a question-authoring layer: three static decision-pattern prompt templates and a guide to one decision per question ID, balanced descriptive criteria, inline state framing and distinct framing per role. Node 20 or newer is required.

Question 5

What does the judge tool take and return?

Accepted Answer

Inputs are a state (object, string or array), a questions map, an optional timeout_ms up to 60,000 and an optional model override. It returns {answers, model, usage} on success or a fallback envelope on failure, and in-flight requests cancel if the client disconnects.

Question 6

What are the three question types?

Accepted Answer

noul omits criteria and returns a yes/no probability; choice takes an option-to-description map and returns the winning option name with probabilities and confidence; score takes an ordered array of 2 to 10 level descriptions and returns the best-matching level.

Question 7

What providers and prompt templates does it offer?

Accepted Answer

The default provider is typesafe-api, with an opt-in cloudflare-workers-ai provider that needs a Cloudflare token and account id; JUDGE_PROVIDER is read once at startup. It also exposes three static prompt templates, verify-claim, classify and route, plus a guide to one decision per question ID and balanced criteria.

Decisions Judge MCP

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community