Skip to main content
Glama

Jev AI MCP Server

Use Jev AI from Claude Code, Cursor and other MCP clients. Classify requests, score content, check conditions, assess proposed actions, answer multiple typed questions, or reuse a saved judge through the Jev AI API.

This is an independent integration for jev-ai.pro, not the official TypeSafe AI MCP server. A TypeSafe API key will not work here.

Quick start

Requires Node.js 20+ and a Jev AI API key. Sign in at jev-ai.pro/jev-api and create a key. Account credits are required for evaluation.

Run directly from GitHub; no build step or npm publication is required:

npx -y github:weidacn/jev-ai-mcp#v1.0.1

Supply the key through your MCP client's environment configuration. Never commit a real key to source control.

Claude Code

claude mcp add jev-ai -e JEV_AI_API_KEY=YOUR_JEV_AI_KEY -- npx -y github:weidacn/jev-ai-mcp#v1.0.1

Cursor / Claude Desktop / generic MCP configuration

{
  "mcpServers": {
    "jev-ai": {
      "command": "npx",
      "args": ["-y", "github:weidacn/jev-ai-mcp#v1.0.1"],
      "env": { "JEV_AI_API_KEY": "YOUR_JEV_AI_KEY" }
    }
  }
}

Use your client's secret storage where available. The server uses stdio. It does not listen on a network port. It can start and list tools without a key; calling tools requires one.

Related MCP server: openrouter-jev-mcp

Tools

Tool

Purpose

jev_classify

Select one of 2–400 labelled options; return confidence and probabilities.

jev_score

Rate input on 2–20 ordered levels; return a fractional zero-based score.

jev_check

Evaluate a yes/no question; return a probability from 0 to 1.

jev_gate

Assess a proposed action; return an advisory recommendation and risk score.

jev_decide

Ask 1–64 typed questions about the same input in one API call.

jev_saved_judge

Run a judge saved in your account; optionally pin its revision.

Every successful tool returns the API response: model, answers, and usage. Single-question tools use answers.result. jev_gate uses answers.recommendation and answers.risk. jev_decide preserves your question names.

Route a support request

Call jev_classify with:

{
  "state": "My payment failed. Can you help?",
  "instructions": "Which team should handle this request?",
  "options": {
    "billing": "Payments, invoices and refunds",
    "technical": "Bugs and outages",
    "sales": "Product and pricing questions"
  }
}

Evaluate several questions together

Call jev_decide with:

{
  "state": "I was charged twice and need this resolved today.",
  "questions": {
    "urgent": { "type": "noul", "instructions": "Does this need urgent support?" },
    "priority": {
      "type": "score",
      "instructions": "How urgent is this request?",
      "criteria": ["Routine", "Today", "Immediate"]
    }
  }
}

Reuse a saved judge

Create a judge on Jev AI, then call jev_saved_judge with its ID and new input:

{ "judgeId": "YOUR_JUDGE_ID", "state": "New text to evaluate", "revision": 1 }

An outdated revision produces a conflict rather than silently using changed rules.

Configuration

Environment variable

Required

Default

JEV_AI_API_KEY

For tool calls

Your key from jev-ai.pro

JEV_AI_MODEL

No

jev-latest

Requests go to https://jev-ai.pro/api/v1/systemone using Bearer authentication. There is no configurable alternate host, so a client cannot redirect your key to a different provider. HTTP redirects are rejected.

Usage, privacy and limits

  • Tool calls transmit the supplied input and instructions to Jev AI and its model providers. Send only data you are authorized to process; consult the site's current privacy policy and API documentation.

  • Calls consume credits. Purchased usage is metered by input tokens; promotional credits are spent per request. See current pricing and account usage for billing details.

  • The API request body must fit within 256,000 UTF-8 bytes. Model context limits still apply. Account and service capacity limits can return HTTP 429.

  • A failed request returns an MCP tool error. HTTP 401 indicates a key problem; 402 indicates insufficient credits or a billing hold; 422 indicates invalid input; 429 indicates rate/capacity limits.

  • Requests time out after 75 seconds and are never automatically retried. A network failure can occur after the service has processed a request; check your account usage before retrying.

  • Decisions are probabilistic. jev_gate provides advice only, never runs a command, and must not replace your application's authorization rules.

  • The MCP process does not write inputs or keys to disk or log API response bodies. Jev AI applies its own service-side data policies.

Develop and test

npm ci
npm test
npm start

Tests exercise the MCP protocol with the official SDK and a mocked HTTP boundary; no paid model calls are required.

MIT licensed. Website: Jev AI · API: Jev AI API

Available Tools

6 tools
jev_checkCheck a conditionA

Evaluate a yes/no condition. answers.result.noul is the probability of yes from 0 to 1, not a guaranteed truth or enforcement decision. Use jev_decide for multiple questions. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: state (required), instructions (required).

jev_classifyClassify inputA

Choose one labelled option for routing, categorization or intent detection. Returns choice, confidence and probabilities in answers.result. Use jev_score for ordered levels. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: state (required), options (required), instructions (required).

jev_decideAnswer multiple typed questionsA

Evaluate 1–64 named questions in one API request. Each type is noul (yes/no), choice (label map) or score (ordered levels). Returns answers keyed by your question names. Prefer this for several decisions on the same input. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: state (required), questions (required).

jev_gateAssess an actionA

Assess an action before execution. Returns an advisory allow/confirm/block classification in answers.recommendation and a zero-based risk score in answers.risk. Does not execute or enforce actions; retain your own permission policy. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: action (required), context.

jev_saved_judgeRun a saved judgeA

Evaluate new input with a judge saved in your Jev AI account. Does not change the judge. Optional revision detects changed rules with HTTP 409. Use jev_decide for inline rules. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: state (required), judgeId (required), revision.

jev_scoreScore inputA

Rate input on an ordered scale of 2–20 levels. Returns a possibly fractional zero-based score, legend and probabilities in answers.result. Use jev_check for yes/no. Requires JEV_AI_API_KEY from https://jev-ai.pro/jev-api. Sends input to Jev AI and consumes account credits; rate limits apply. Returns model, answers and usage. No automatic retries.

Inputs: state (required), levels (required), instructions (required).

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.5/5.0

    Scored across 6 tools

    Disambiguation4/5

    Each tool has a distinct purpose: classification, scoring, yes/no checking, gating, multi-question decisions, and saved judge evaluation. The descriptions cross-reference each other (e.g., 'Use jev_score for ordered levels'), which helps disambiguate, though jev_gate and jev_check could be confused in some safety contexts.

    Naming Consistency5/5

    All tools follow a consistent 'jev_' prefix with a clear verb or noun indicating the operation: classify, score, check, gate, decide, saved_judge. The naming pattern is uniform and predictable.

    Tool Count5/5

    Six tools is well-scoped for a decision/evaluation API. Each tool covers a distinct mode of interaction (single vs. batch, inline vs. saved judge), and none feel redundant or excessive.

    Completeness4/5

    The set covers the core decision types: classification, scoring, yes/no, gating, batch decisions, and saved judges. A minor gap is the lack of a tool to create/manage saved judges, but the server's stated purpose is evaluation, not judge management.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      B
      quality
      C
      maintenance
      Enables AI agent skills to route state-evaluation requests to TypeSafe AI's Jev System One model, supporting typed questions, choice classification, binary probabilities, and rubric scoring with calibrated confidence.
      4
      2
      MIT
    • A
      license
      B
      quality
      C
      maintenance
      Enables coding agents to obtain probabilistic decisions from Jev AI via OpenRouter for classification, scoring, and validation, with tools like jev_check, jev_classify, jev_score, and jev_evaluate.
      5
      3
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Enables coding or reasoning agents to request structured judgments from TypeSafe's Jev model at decision points, including choices, scores, claim verification, and code reviews, with probabilities and confidence returned as data.
      5
      13 npm
      MIT
    • A
      license
      B
      quality
      C
      maintenance
      Enables 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.
      5
      2
      AGPL 3.0