askjev
Ask Jev, a fast non-generative judgment model, one or more plain-language questions about material you already have, and get calibrated probabilities back instead of a generated sub-agent response.
Ask yes/no questions and receive the probability of "yes".
Ask scale/rating questions; provide ordered options for a custom rubric or let a built-in rubric (severity, likelihood, sentiment, etc.) be chosen automatically.
Ask choice questions among named alternatives and get probabilities per option.
Batch multiple questions about the same state in one round trip.
Get per-answer probabilities, confidence, routing metadata, and optional token usage/model info.
Handle individual bad questions gracefully: one error does not fail the rest of the batch.
Works locally via stdio or hosted over HTTP, with support for several MCP clients (Claude Code, Claude Desktop, Cursor, Codex).
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@askjev Based on this customer email, should we offer a refund?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
askjev
Unofficial MCP server for Jev, Typesafe AI's System One model. Not affiliated with Typesafe AI.
Give your agent a fast second opinion. It asks a plain question about material it already has, Jev works out whether that is a yes/no, a scale, or a choice, and answers with calibrated probabilities. No generative model in the loop, one round trip, a fraction of the cost and latency of a sub-agent.
Demo
An agent planning a workflow, asking Jev one question at a time. Every number is real.

Related MCP server: jev-mcp
Install
You need a Typesafe API key. The server runs in one of two places; every client below supports at least one.
Hosted.
https://jev.zacca.dev/mcpruns this repository on Cloudflare Workers. It keeps nothing: every call builds a Typesafe client from the key you send and forwards the question. Send the key in thex-api-keyheader, or as a bearer token if your client only has that field. The same server is listed on Smithery as pzacca/askjev.Local.
npx -y askjevruns it on your machine over stdio. Requires Node 22+ and the key inTYPESAFE_API_KEY.
Supported clients
Client | Local (stdio) | Hosted (HTTP) |
yes | yes | |
yes | no, custom connectors cannot send an API key header | |
yes | yes | |
yes | yes, as a bearer token |
Claude Code and Codex were exercised end to end on both transports. Cursor follows its documented configuration format.
Claude Code
# hosted
claude mcp add --transport http askjev https://jev.zacca.dev/mcp --header "x-api-key: your-key"
# local
claude mcp add askjev -e TYPESAFE_API_KEY=your-key -- npx -y askjevClaude Desktop
Open Settings, Developer, Edit Config. The file is
~/Library/Application Support/Claude/claude_desktop_config.json on macOS and
%APPDATA%\Claude\claude_desktop_config.json on Windows.
{
"mcpServers": {
"askjev": {
"command": "npx",
"args": ["-y", "askjev"],
"env": { "TYPESAFE_API_KEY": "your-key" }
}
}
}Restart Claude Desktop after saving. On Windows some hosts cannot launch npx directly;
use "command": "cmd" with "args": ["/c", "npx", "-y", "askjev"].
Cursor
Add to ~/.cursor/mcp.json for every project, or .cursor/mcp.json for one project.
{
"mcpServers": {
"askjev": {
"url": "https://jev.zacca.dev/mcp",
"headers": { "x-api-key": "your-key" }
}
}
}For a local server use the same command, args and env block as Claude Desktop.
Codex
# hosted: Codex reads the key from an environment variable and sends it as a bearer token
export TYPESAFE_API_KEY=your-key
codex mcp add askjev --url https://jev.zacca.dev/mcp --bearer-token-env-var TYPESAFE_API_KEY
# local
codex mcp add askjev --env TYPESAFE_API_KEY=your-key -- npx -y askjevThe tool
One tool, ask. It takes the material and a list of free-text questions. Options are only
needed when the question has named alternatives.
Input
{
"state": "...", // string | object | array: the material to judge
"questions": [
{ "question": "Is this a bug report?" },
{ "question": "How severe is it?" },
{ "question": "Which team owns it?", "options": ["billing", "platform", "mobile"] },
{ "question": "How urgent is it?", "options": ["can wait", "this week", "today"] }
]
}questionis free text. Jev decides whether it is a yes/no question, a scale, or a choice between the given options.optionsis optional. Pass it when the question has named alternatives. Order matters when the options form a scale.There is no way to force the question type. The whole point is that the agent does not have to think about it.
Output
One entry per question, in input order.
{
"answers": [
{
"kind": "noul",
"answer": 0.93, // probability of "yes"
"probabilities": { "yes": 0.93, "no": 0.07 },
"routing": { "kind": "noul", "confidence": 0.98 }
},
{
"kind": "score",
"answer": 3.4, // expected level, may be fractional
"legend": { "0": "trivial", "1": "minor", "2": "moderate", "3": "major", "4": "critical" },
"probabilities": { "0": 0.01, "1": 0.04, "2": 0.15, "3": 0.5, "4": 0.3 },
"confidence": 0.71,
"rubric": "severity",
"note": "No options were given, so the built-in 'severity' rubric was used. Pass options for a rubric tailored to your question.",
"routing": { "kind": "score", "confidence": 0.9, "rubric": { "name": "severity", "confidence": 0.84 } }
},
{
"kind": "choice",
"answer": "platform",
"probabilities": { "billing": 0.05, "platform": 0.88, "mobile": 0.07 },
"confidence": 0.88,
"routing": { "kind": "choice", "confidence": 0.97 }
},
{
"kind": "score",
"answer": 1.8,
"legend": { "0": "can wait", "1": "this week", "2": "today" },
"probabilities": { "0": 0.1, "1": 0.2, "2": 0.7 },
"confidence": 0.7,
"routing": { "kind": "score", "confidence": 0.79 }
}
]
}A question that cannot be answered does not fail the batch. Its entry has kind: "error"
and a message saying what to change, and every other question is still answered:
{
"kind": "error",
"message": "\"Which team owns it?\" asks to pick between alternatives, but no options were given. Pass options with the alternatives.",
"routing": { "kind": "choice", "confidence": 0.95 }
}Everything is raw. There is no threshold and no verdict. confidence comes straight from
Jev; yes/no answers have no separate confidence because the probability is the signal.
routing exposes how sure Jev was about the question type, and about the rubric when one
was picked, so a misroute is visible rather than silent.
What else comes back is the operator's choice, not the agent's. ASKJEV_INCLUDE is a
comma-separated list of the optional parts: routing (the default), usage (tokens summed
over every Jev call), and model. Set it to usage,routing to add token counts, or to an
empty string to get nothing but the answers. The tool description tells the agent what the
server was configured to emit.
How it works
flowchart TD
A[ask: state + questions] --> B[Call 1: route<br/>one Jev call, one choice question per input question]
B --> C{per question}
C -->|options given| D{choice or score?}
C -->|no options| E{noul, score, or<br/>choice-without-options?}
D -->|choice| F[choice with options as labels]
D -->|score| G[score with options as ordered rubric]
E -->|noul| H[noul]
E -->|score| I[Call 2: pick rubric<br/>one Jev call, one choice question per such input question]
E -->|choice| X[error: this question needs options]
I --> J[score with built-in rubric]
F --> K[Call 3: answer<br/>one Jev call with every typed question]
G --> K
H --> K
J --> K
K --> L[shape answers + routing metadata + summed usage]Two Jev calls in the common case, three when at least one question is a scale without
options. The number of input questions does not change the number of calls: each step
batches every question that needs it into a single systemone request.
Routing table
| Router decides between | Becomes |
2+ items | choice, score |
|
absent | noul, score, choice |
|
1 item | rejected by schema validation | never reaches the router |
The router's state is the question text itself (an object keyed by question when
batching), and its instructions ask which kind of question that text is. The routing
criteria are described in plain language so Jev discriminates on intent, not on keywords.
Built-in rubrics
Used only for scale questions that arrive without options. A second Jev call picks the closest one. All have five levels so Jev has room to discriminate without the levels blurring together.
Name | Levels, lowest to highest |
| not at all, slightly, moderately, very, extremely |
| poor, below average, acceptable, good, excellent |
| trivial, minor, moderate, major, critical |
| very unlikely, unlikely, uncertain, likely, very likely |
| very negative, negative, neutral, positive, very positive |
| never, rarely, sometimes, often, always |
| strongly disagree, disagree, neutral, agree, strongly agree |
The list is fixed in code. Configurable rubrics are a possible later addition.
Errors
Situation | Behaviour |
Input fails schema validation | Tool error with the Zod message. Jev is not called. |
Question routed to | That entry becomes |
Jev returns 401 | Tool error: API key missing or invalid, with the env var name. |
Jev returns 422, 429 after retries, 5xx | Tool error with status, Jev's message, and the request id when present. |
Network or timeout after retries | Tool error with the SDK's message. |
Errors that affect the whole call are returned as MCP tool results with isError: true,
not as protocol errors, so the agent sees the message and can recover. Errors that affect
one question are entries in answers, so one bad question never costs the agent the
others.
Configuration
Local, only what the Typesafe SDK already reads from the environment:
Variable | Meaning |
| Required. |
| Optional, defaults to |
| Optional, for proxies and the smoke test stub. |
| Optional. SDK logs go to stderr. |
Plus one setting of the server's own:
Variable | Meaning |
| Optional parts of every result, comma-separated: |
stdout carries MCP protocol messages only; every diagnostic goes to stderr.
Hosted, the key travels per request in the x-api-key header (or Authorization: Bearer). TYPESAFE_BASE_URL, TYPESAFE_DEFAULT_MODEL and ASKJEV_INCLUDE can be set as
Worker vars and apply to every caller. The public instance at jev.zacca.dev runs the
defaults.
Evaluation
The router is the part that can be wrong in a way unit tests cannot catch. A versioned dataset of human-labelled questions runs against the live API weekly and on demand; the first run scored 42/42 on routing and 16/16 on rubric selection. See docs/EVAL.md. Design decisions and rejected alternatives are in docs/ARCHITECTURE.md.
Development
npm ci
npm run lint
npm run typecheck
npm testThe hosted variant is src/worker.ts, bundled and deployed by wrangler from
wrangler.jsonc. npm run dev serves it on localhost. Every push to main that passes
CI is deployed to jev.zacca.dev by the deploy job in ci.yml; npm run deploy does
the same by hand (needs wrangler login).
Releases are tag-driven. Bump the version in package.json, commit, then:
git tag v0.2.0 && git push origin main v0.2.0release.yml checks that the tag matches package.json, runs the full test suite,
publishes to npm with provenance, and creates the GitHub release with generated notes.
The Smithery listing points at that domain and takes its configuration form from
smithery.schema.json. Smithery stores a snapshot of the tool taken at publish time, so
the same deploy job republishes the listing after every deploy to keep it current.
License
MIT
Available Tools
1 toolaskAsk JevARead-only
Ask Jev, a fast non-generative judgment model, one or more questions about some material you already have. Use it for a quick second opinion, a classification, a rating, or a yes/no check instead of reasoning it out yourself or spawning a sub-agent.
Write each question in plain language. Jev decides whether it is a yes/no question, a scale, or a choice:
Yes/no: "Does this ticket ask for a refund?" You get the probability of yes.
Choice: pass "options" with the named alternatives. "Which team owns this?" with options ["billing", "platform"].
Scale: pass "options" as ordered levels, lowest first. "How urgent is this?" with options ["can wait", "this week", "today"]. Without options, a built-in rubric is picked for you:
intensity: not at all < slightly < moderately < very < extremely
quality: poor < below average < acceptable < good < excellent
severity: trivial < minor < moderate < major < critical
likelihood: very unlikely < unlikely < uncertain < likely < very likely
sentiment: very negative < negative < neutral < positive < very positive
frequency: never < rarely < sometimes < often < always
agreement: strongly disagree < disagree < neutral < agree < strongly agree
Every answer carries probabilities and Jev's confidence, plus "routing" showing how the question was interpreted and how sure Jev was about that. Read the confidence: a low value means the material does not settle the question, so add context or decide another way.
Requires a Typesafe API key. Set TYPESAFE_API_KEY in the environment of the askjev process.
Inputs: state (required), questions (required). Returns: model, usage, answers
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 1 tool
With only one tool, there is no possibility of confusion or misselection. The tool's purpose is clearly defined and unique.
A single tool name 'ask' is trivially consistent; there is no pattern to break or mix. The name accurately reflects its function.
A single tool is on the borderline of being thin, but it is purpose-built for a narrow domain (querying a judgment model). The count is acceptable but could feel minimal for broader expectations.
The tool fully covers the intended domain: asking yes/no, choice, or scale questions, with built-in rubrics and confidence reporting. No obvious gaps in its stated purpose.
Maintenance
Related MCP Connectors
Calibrated judgments for text: yes/no probabilities, picks from your options, or scores.
Typed-question decisions from Jev: classify, moderate, route models and guard tool calls.
Grade essays, check resumes, and ask typed yes/no, choice and rubric questions, powered by Jev.
Jev-powered decisions, web search, PDF/web to Markdown, summarize. From $0.001, no API key.
Related MCP Servers
- AlicenseAqualityCmaintenanceEnables frontier coding agents to delegate routine probabilistic judgments to TypeSafe Jev, providing calibrated triage signals for failures, attempts, completion, context ranking, findings, risk, and generic evidence-grounded questions.7MIT
- AlicenseBqualityCmaintenanceEnables AI agents to obtain typed judgments from TypeSafe's Jev System One models, including yes/no probabilities, multiple-choice selections with distributions, and rubric-based scores, directly usable in code.52AGPL 3.0
- AlicenseAqualityBmaintenanceEnables AI coding agents to offload yes/no, multiple-choice, and scoring questions to TypeSafe's Jev, returning compact confidence-scored answers to save tokens and improve speed.1MIT
- AlicenseAqualityCmaintenanceEnables agents to query the TypeSafe Jev decision model for typed answers—yes/no, choice, and score questions—with calibrated probabilities, batched in a single request under a local context-budget guard.6MIT