jev-mcp
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., "@jev-mcp Should we approve this loan? Give me a yes/no probability."
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.
jev-mcp
MCP server wrapping TypeSafe's Jev System One models. Instead of prompt-and-parse, agents get typed judgments they can use directly in code: yes/no probabilities, multiple-choice selections with distributions, and rubric-based scores.
Tools
Tool | What it does | Returns |
| Evaluate |
|
| Yes/no judgment |
|
| Pick one option from |
|
| Rate along 2–10 ordered level descriptions |
|
| List models your key can use |
|
Every tool takes state (string, object, or array — the content to judge),
an optional model (default "jev-latest"), and an optional endpoint
("direct" or "zen", default "direct").
Related MCP server: Jev MCP Proxy
Prerequisites
Python 3.10+ and
uv(or plainpip)A TypeSafe API key →
TYPESAFE_API_KEYOptional: an OpenCode Zen key →
OPENCODE_API_KEY, only if you wantendpoint="zen"(modelsjev-1.13/jev-1.13-free)
Quickstart
git clone https://github.com/rajasekharponakala/jev-mcp.git
cd jev-mcp
cp .env.example .env # then put your key in .env (never commit it)
uv run --with fastmcp --with httpx server.pyOr with an installed environment:
pip install -e .
export TYPESAFE_API_KEY=...
jev-mcpClient configuration
OpenCode (~/.config/opencode/opencode.json)
{
"mcp": {
"servers": {
"jev": {
"type": "local",
"command": ["uv", "run", "--with", "fastmcp", "--with", "httpx", "server.py"],
"cwd": "/path/to/jev-mcp",
"environment": {
"TYPESAFE_API_KEY": "{env:TYPESAFE_API_KEY}"
}
}
}
}
}Claude Code
claude mcp add jev -e TYPESAFE_API_KEY=... -- uv run --with fastmcp --with httpx /path/to/jev-mcp/server.pyCodex CLI (~/.codex/config.toml)
codex mcp add jev --env TYPESAFE_API_KEY=... -- uv run --with fastmcp --with httpx /path/to/jev-mcp/server.pyor manually:
[mcp_servers.jev]
command = "uv"
args = ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"]
env = { TYPESAFE_API_KEY = "..." }Gemini CLI (~/.gemini/settings.json)
{
"mcpServers": {
"jev": {
"command": "uv",
"args": ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"],
"env": { "TYPESAFE_API_KEY": "..." }
}
}
}Cursor (~/.cursor/mcp.json)
{
"mcpServers": {
"jev": {
"command": "uv",
"args": ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"],
"env": { "TYPESAFE_API_KEY": "..." }
}
}
}Windsurf (~/.codeium/windsurf/mcp_config.json)
{
"mcpServers": {
"jev": {
"command": "uv",
"args": ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"],
"env": { "TYPESAFE_API_KEY": "..." }
}
}
}Cline (~/.cline/mcp.json, or cline mcp wizard)
{
"mcpServers": {
"jev": {
"command": "uv",
"args": ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"],
"env": { "TYPESAFE_API_KEY": "..." },
"disabled": false
}
}
}Goose (~/.config/goose/config.yaml)
goose session --with-extension "jev:TYPESAFE_API_KEY=... uv run --with fastmcp --with httpx /path/to/jev-mcp/server.py"or persist in config:
extensions:
jev:
name: Jev
cmd: uv
args: ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"]
enabled: true
envs: { "TYPESAFE_API_KEY": "..." }
type: stdio
timeout: 300Claude Desktop / VS Code / other MCP JSON clients
{
"mcpServers": {
"jev": {
"command": "uv",
"args": ["run", "--with", "fastmcp", "--with", "httpx", "/path/to/jev-mcp/server.py"],
"env": { "TYPESAFE_API_KEY": "..." }
}
}
}(VS Code: same shape in ~/.vscode/mcp.json.)
FastMCP one-liner
The server object is named mcp, so FastMCP can install it directly:
fastmcp install claude-code /path/to/jev-mcp/server.pythen add TYPESAFE_API_KEY via that client's config.
Example
// tool call: jev_choice
{
"state": "Help! My payouts have been failing for 3 days.",
"instructions": "Which team should handle this?",
"criteria": {
"billing": "Payments, invoicing, refunds",
"technical": "Bugs, outages, integrations",
"sales": "Pricing, upgrades, new accounts"
}
}
// → {"choice": "billing", "probabilities": {"billing": 0.88, ...}, "confidence": 0.81}See the TypeSafe docs (source of truth) and the API reference for question design: one narrow judgment per question, batch independent questions, keep policy and thresholds in your code.
Development
uv run --with fastmcp --with httpx python -c "
import asyncio, server
print(sorted(t.name for t in asyncio.run(server.mcp.list_tools())))
"License
AGPL-3.0-only. Copyright (c) 2026 rajasekharponakala.
Available Tools
5 toolsjev_choiceJev ChoiceB
Pick one option. criteria: {"option": "rubric..."}. Returns choice + probabilities + confidence.
Inputs: state (required), criteria (required), instructions (required), model, +1 more. Returns: object
jev_evaluateJev EvaluateA
Evaluate state against typed questions. Returns {model, answers, usage}.
questions: {"id": {"type": "noul"|"choice"|"score", "instructions": ..., "criteria": {...} or [...]}}. Batch independent questions in one call; they run in parallel. endpoint: "direct" (TYPESAFE_API_KEY) or "zen".
Inputs: state (required), questions (required), model, endpoint. Returns: object
jev_modelsJev ModelsA
List models your key can use (GET /v1/models on direct; static list on zen).
Inputs: endpoint. Returns: object
jev_noulJev NoulC
Yes/no judgment. Returns {type, noul (0=no..1=yes)}. Near 0.5 = uncertain.
Inputs: state (required), instructions (required), model, endpoint, +2 more. Returns: object
jev_scoreJev ScoreC
Rate along ordered levels (2-10 concrete descriptions). Returns score + legend + probabilities.
Inputs: state (required), criteria (required), instructions (required), model, +1 more. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
jev_evaluate subsumes the other evaluation tools, creating potential overlap. The descriptions clarify that evaluate is for batched/mixed types while the specialized tools handle single questions, but an agent could still be unsure which to use for a one-off noul or choice query.
All tools share the 'jev_' prefix in lowercase snake_case, making them easily recognizable as a family. The names mix verbs and nouns (evaluate, models vs noul, choice, score), but the uniform prefix and short, predictable tokens keep the pattern strong.
Five tools is a well-scoped size for an evaluation API wrapper. It includes a general evaluator, three specialized question types, and a model discovery utility without unnecessary bloat.
The toolset covers all declared question types (noul, choice, score) and provides the necessary model listing capability. It lacks history or configuration management, but those are not implied by the server's focused purpose.
Related MCP Connectors
Grade essays, check resumes, and ask typed yes/no, choice and rubric questions, powered by Jev.
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.
Typed AI decisions for agents: classify, route, screen evidence, score and batch-label text.
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 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.42MIT
- AlicenseAqualityCmaintenanceEnables 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.513 npmMIT
- FlicenseNot gradedqualityBmaintenanceEnables Claude Code agents to perform structured decision-making through TypeSafe Jev, supporting yes/no checks, route selection, scoring, and batched multi-question judgments with confidence values.1-