Question 1

Why batch every question into one call instead of asking separately?

Accepted Answer

Because TypeSafe bills per input token, so sending one state per question pays for that state repeatedly. `ask` takes a `questions` array over a single state and returns answers keyed by your own ids, and the README notes that batching is barely slower than a single question. Mixed types are the case it exists for: a choice, a noul and a score over the same state in one request.

Question 2

What does the `confidence` field actually mean, and can I threshold on it?

Accepted Answer

The README treats it as a separate axis from the answer: the answer tells you what, confidence tells you whether to act. It comes back on choice and score answers, and the recommended pattern is to gate on it, acting automatically above your threshold and routing to a human below. It is not present on noul answers, where the returned value is itself the probability of yes.

Question 3

What happens when I get an argument wrong or the API is down?

Accepted Answer

Two different channels. Invalid arguments are rejected by the input schema before any request is made and surface as a protocol error, not a tool result. API failures come back as `isError` tool results carrying the status and body. Retries are the SDK's RetryPolicy defaults: 408, 429 and 500-599 plus connection and timeout failures, two retries after the first attempt, 500ms backoff doubling to 5s with 25% jitter, honoring Retry-After up to 60s, with a 10s per-attempt timeout and no total budget.

Question 4

Are the benchmark results in the README convincing?

Accepted Answer

Not as evidence that the model separates quality. The bench compares the author's own content scorer against Jev on two judgment dimensions and reports analyzer AUC 0.52, Jev 0.52, hybrid 0.49 over 30 marketing documents, with the author stating no scorer separated published from draft. What it did confirm is saturation: the analyzer's Specificity has a standard deviation of 3.1 around a mean of 99, versus 15.8 for Jev's.

Question 5

What do the two tools return beyond the answers?

Accepted Answer

ask takes a questions array over one state and returns answers keyed by your ids; list_models reports which models the account can use. Either way the response also carries the resolving model and token usage, and ask takes an optional model to override the server default for that call. Because it is built on TypeSafe's official @typesafe-ai/sdk, the wire format, retries and error classes are upstream's rather than reimplemented here.

Question 6

How do I install it, and which MCP era does it speak?

Accepted Answer

Node 20 or newer, the v2 SDK's floor, then npm install and npm run build. It is built on SDK v2 with @modelcontextprotocol/server and serves protocol revision 2026-07-28 through serveStdio. Clients that still open with the 2025 initialize handshake are served too, because the entry point handles both eras, so older hosts keep working. Configuration is TYPESAFE_API_KEY, plus optional TYPESAFE_BASE_URL, default https://api.typesafe.ai, and TYPESAFE_DEFAULT_MODEL, default jev-latest.

Question 7

How do I sanity-check it against real notes?

Accepted Answer

scripts/vault-test.mjs runs real notes from the vault's 40-RESOURCES directory through the server over stdio, the same path a client takes, so it exercises the MCP layer rather than just the HTTP client. Each note becomes one batched ask with three questions: a choice for the primary topic drawn from the vault's own tag taxonomy, a noul for whether it is actionable, and a score for how deep. The hand-written tags are the ground truth, and only the title and TL;DR are sent, about 700 input tokens per note.

AskJev-MCP

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community