Ask TypeSafe's Jev for typed, calibrated judgments from
your AI assistant: how likely a yes/no statement is to hold, which of your
options fits, or where something sits on your own scale. Put the same
questions to one state or to up to 1000 rows at once, so the assistant can
triage postings, transactions or emails without reading each one. Every tool
is a read: nothing changes anywhere, and no tool is money.
You need a TypeSafe account and an API key from its console. Calls are
billed by input tokens, a fraction of a cent each, and Jev answers in about
half a second.
What you can ask
- "Go through the 300 postings in postings.jsonl and list the remote senior
roles, with how sure Jev is of each."
- "Is this support email urgent, and which team should take it: billing,
technical or sales?"
- "Sort last month's 600 card transactions into groceries, transport, software
or other, and count each."
- "Rate these 40 reviews on how likely the customer is to leave, and show me
the ten most at risk."
- "Which of these passages answer 'how do I reset my password'? Flag the ones
Jev is unsure about for me to read."
Install
[Install in VS Code] [Install in Cursor] [Download for Claude Desktop]
With Claude Code:
Claude Desktop (Settings, Developer, Edit Config) and most other clients read the same command as JSON:
Each runs the published package on Node 22.19 or newer. VS Code and the Claude Desktop bundle ask for TYPESAFE_API_KEY_FILE as they add the server. Cursor reads it from your environment instead (${env:TYPESAFE_API_KEY_FILE}), so set it before you start Cursor. Configuration says what it holds.
From a clone
In a JSON configuration, the command is node and its one argument the full path of servers/typesafe/dist/index.js. To share one running copy between several clients over HTTP, see docs/deploying.md.
Configuration
Without a key the server still starts and lists its tools, and each call
answers a config error naming the variable. ask_batch reads an
items_file from MCP_FILE_ROOTS; that and the other variables every server
shares are in docs/configuration.md.
Getting an API key
- Sign in at https://console.typesafe.ai and open Keys.
- Create a key and write it to a file only you (or the server's user) can
read.
- Point
TYPESAFE_API_KEY_FILE at that file.
Give each deployment its own key, so you can revoke one alone and tell them
apart on the console's usage page. Only a key from TypeSafe's own console
works; a reseller's key works only through the reseller's gateway.
First run
Ask your assistant to run TypeSafe's health check. It answers with the API
it talks to, the default model, and the models your key reaches. A missing key
shows up here as config, a refused one as auth, before any batch starts.
Tools
read changes nothing; every tool here is a read (billing by input tokens is
metering, not a payment).
A question is a noul (the probability that a yes/no statement holds), a
choice (one of 2 to 255 options, with a probability for each) or a score
(the state placed on 2 to 10 levels you order from lowest to highest).
Troubleshooting
config. No key: set TYPESAFE_API_KEY_FILE to a file holding a key from
console.typesafe.ai/keys, or TYPESAFE_API_KEY for a local run.
auth. TypeSafe refused the key. Check it still exists at
console.typesafe.ai/keys, and replace the file if it was revoked.
input. The question cannot be answered as asked; the message names the
field (a choice needs at least two options, a score 2 to 10 levels). Above 20
rows, every row needs an id.
rate_limited. TypeSafe is still throttling or overloaded after the
server's own retries. Wait a minute, or send the batch with a lower
concurrency.
api. Any other refusal, a 5xx that outlived its retries, or no connection
(status 0). Retry once; if it persists, run health and read the server's log.
drift. TypeSafe answered in a shape this server does not know (its SDK's
types may have moved on). Please report it.
How it works
The network. An evaluation has no side effect, so every request
(POST /v1/systemone, GET /v1/models) may repeat: a 5xx or a dropped
connection is tried again, and a 429 or 529 ("overloaded") waits as a rate
limit, as TypeSafe's SDK does. A call times out after a minute. Every answer
is checked against the SDK's types (@typesafe-ai/sdk, types only, in
src/domain/schema.ts).
Batches. A batch goes on past a failed row and reports it, but stops at a
failure that would fail every row (no key, a refused key), so a thousand rows
do not each earn a 401. Answers join their rows by id, required above 20 rows:
a position is only safe to join on while you can check the rows by eye.
Rows from a file. Instead of typing rows into items, pass items_file:
JSONL with one {"id", "state"} object per line, or a JSON array of them, by
path inside MCP_FILE_ROOTS (or as content_base64 with its sha256, from a
script). The rows never pass through a model, so they cannot change on the
way; a malformed row fails the call with its line number before anything is
sent. The reply names the file's sha256 and size.
Verdicts. With thresholds for a question, each row's answer gets act
(sure enough to act on), review (a person should look) or abstain (do not
use it), from the answer's certainty: for a noul |p - 0.5| x 2 (p 0.95 and
p 0.05 are both 0.9), for a choice or a score the confidence TypeSafe returns.
The certainty is rounded down to three places, the number the row shows, and
must be strictly above act_above to act and above review_above to be
reviewed. Once thresholds are tuned, pin model to a versioned Jev:
jev-latest can move under them.
Choosing rows. where lists only the answered rows meeting every
condition you give per question: true or false for a noul (a yes is
p >= 0.7, a no p <= 0.3, as the summary counts them), labels for a choice,
min and max on the value a row shows, and verdict. top keeps the k
rows with the highest p or score for one question, and counts tallies the
answers of the rows where kept; the summary still covers every row. Every
condition is checked before a request is sent, so a label that is not an
option fails the call, not a thousand rows.
npx tsx scripts/smoke.ts drives the built server against the real API, for
a few thousand input tokens: health, one ask per question type, a
five-row batch saved to the outbox, and one input failure that never reaches
TypeSafe.
Limitations
- Judgments only. Jev writes no text and does no multi-step reasoning.
Keep rules, lookups and arithmetic in code, and threshold the probabilities
(
thresholds does it per question).
where joins conditions with "and" only. For "or", read the rows from a
saved JSONL, or ask a question that states the alternatives.- At most 1000 rows a call; split a larger set.
- A batch's reply lists every row unless
list says otherwise. For a
large batch, pass list: "none" or "failed" with save, and read the
JSONL.
Other servers for TypeSafe
Searched 2026-10-15. TypeSafe publishes none: its GitHub organisation has the
SDKs, agent skills and an n8n node. Two community servers are on npm:
typesafe-mcp: stdio under Bun, the key in an environment variable,
batches over files matched by a glob, with an act, review or abstain verdict
per answer.@y0usaf/typesafe-mcp: stdio, one evaluate tool.
This one exists for what neither offers: it also runs over HTTP as a service
of its own, with its key in a file the calling client never reads; it saves a
batch's full answers as JSONL in an outbox and names the file's SHA-256; and
it follows the conventions every server in this repository shares. Its
verdicts and its where, top and counts follow typesafe-mcp's, with
thresholds set per question rather than per call.
