How is this different from calling the HTTP API from my own code?

You get five ready tools instead of one raw endpoint: a batched evaluate, the three primitives as separate tools, and a model lister that reports what your key can use. State is accepted as a string, object or array, the model defaults to jev-latest and the endpoint can be switched to zen. That is convenience, not new capability, so if you already have an MCP client wrapper the win is the config snippets.

Does it validate my questions, or fix the prose-parsing problem?

Only the latter. The server forwards your typed questions and returns the typed answers; it does not check whether a question is well formed, does not pick thresholds for you and does not retry. The README defers to TypeSafe's llms.txt and API reference as the source of truth for question design and recommends keeping policy and thresholds in your own code.

Is AGPL-3.0 a problem if I want this inside a closed product?

That is your call to make with your own counsel, but note the license is AGPL-3.0-only rather than MIT or Apache, so the network-use copyleft obligations are the part to think about. Running it as a separate local stdio process started by the agent is a different arrangement from linking it into your code, and the README does not discuss licensing boundaries at all.

What does it cost per call, and is there any rate limiting?

The README gives no cost figures, no latency figures and no rate limiting or retry logic, so none of that is handled for you. The only batching claim is that questions inside one jev_evaluate call run in parallel, and the response includes usage. If a tool call loops, you are the one who has to notice.

What is the zen endpoint option?

Every tool accepts endpoint direct or zen, defaulting to direct. Zen is optional and uses an OpenCode Zen key with the models jev-1.13 or jev-1.13-free.

Which MCP clients have ready configs?

The README gives config for OpenCode, Claude Code, Codex CLI, Gemini CLI, Cursor, Windsurf, Cline, Goose and Claude Desktop or VS Code, plus a FastMCP one-liner for Claude Code.

What do the primitive tools return?

Noul returns one probability where 0 is no and 1 is yes with near 0.5 uncertain; choice returns choice, probabilities and confidence; score returns score, legend, probabilities and confidence; models returns what your key can use.

CommunityJev Agents & MCP

Jev MCP — rajasekharponakala

Python MCP server exposing Jev System One as five typed tools, including a batched evaluate that runs a map of questions in one call, so agents consume probabilities instead of parsing prose.

See highlights

Open github.com

Stars: 2
Language: Python
License: AGPL-3.0
Published by: Community
Author: rajasekharponakala
Last verified: Oct 2026

rajasekharponakala/jev-mcp

User repository on GitHub

View on GitHub

MCP server wrapping TypeSafe's Jev System One models — typed noul/choice/score judgments for AI agents

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

This is one of several Jev listings for the same project. The canonical entry is Jev MCP — jkudish.

About this project

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. 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"). Python 3.10+ and uv (or plain pip) A TypeSafe API key → TYPESAFEAPIKEY.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Ranking Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

jev_evaluate sends a map of typed questions in one call and they run in parallel, returning answers plus usage.
jev_choice returns the option with per-option probabilities and confidence; jev_score adds a legend; jev_noul returns 0 to 1.
Every tool takes state as a string, object or array, an optional model defaulting to jev-latest, and endpoint direct or zen.
Configuration blocks are given for OpenCode, Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, Goose and VS Code.

Watch out

This is a typed pass-through, not a decision layer: question design, thresholds, retries, caching and cost control stay in your code, and the README's own advice is one narrow judgment per question with policy held on your side. It needs TYPESAFE_API_KEY; endpoint zen additionally needs OPENCODE_API_KEY and is limited to jev-1.13 and jev-1.13-free. Every client config example in the README inlines the key as a literal string, so pasting one as-is leaves a real key sitting in a config file. Installation is uv run --with fastmcp --with httpx server.py, with fastmcp and httpx resolved at launch rather than pinned in the repo, or pip install -e . for a jev-mcp command. The license is AGPL-3.0-only, which matters if you would rather not ship copyleft code. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-judge-mcp

MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.