jev-code

Name: jev-code
Author: French Castle

Typed classification, yes/no checks, scoring and ranking for coding agents via TypeSafe's Jev model

By French Castle (individual) TypeScript Local · stdio AI and ML Developer tools

GitHub npm Website

44 popularity score ★ 58 ↓ 702/wk

Connect your client

Local (npm)

Add in one click

Add to Cursor Add to VS Code Add to LM Studio Add to Goose

Opens the app, which asks before it adds the server. Then replace each <NAME> in its settings with your value.

Or add it by command or config

It needs TYPESAFE_API_KEY. Replace each <NAME> with your value.

Claude CodeTerminal

claude mcp add jev-code -e TYPESAFE_API_KEY=<TYPESAFE_API_KEY> -- npx -y @french-castle/jev-code@0.5.0 mcp

Claude Desktopclaude_desktop_config.json

{
  "mcpServers": {
    "jev-code": {
      "command": "npx",
      "args": [
        "-y",
        "@french-castle/jev-code@0.5.0",
        "mcp"
      ],
      "env": {
        "TYPESAFE_API_KEY": "<TYPESAFE_API_KEY>"
      }
    }
  }
}

Cursor~/.cursor/mcp.json

{
  "mcpServers": {
    "jev-code": {
      "command": "npx",
      "args": [
        "-y",
        "@french-castle/jev-code@0.5.0",
        "mcp"
      ],
      "env": {
        "TYPESAFE_API_KEY": "<TYPESAFE_API_KEY>"
      }
    }
  }
}

VS Code.vscode/mcp.json

{
  "servers": {
    "jev-code": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "@french-castle/jev-code@0.5.0",
        "mcp"
      ],
      "env": {
        "TYPESAFE_API_KEY": "<TYPESAFE_API_KEY>"
      }
    }
  }
}

Codex~/.codex/config.toml

[mcp_servers.jev-code]
command = "npx"
args = ["-y", "@french-castle/jev-code@0.5.0", "mcp"]

[mcp_servers.jev-code.env]
TYPESAFE_API_KEY = "<TYPESAFE_API_KEY>"

Gemini CLITerminal

gemini mcp add --scope user -e TYPESAFE_API_KEY=<TYPESAFE_API_KEY> jev-code npx -- -y @french-castle/jev-code@0.5.0 mcp

Settings 10

`TYPESAFE_API_KEY`	Required: Key for TypeSafe or a System One gateway. Alternatives: OPENROUTER_API_KEY, AI_GATEWAY_API_KEY, or OPENAI_API_KEY with JEV_CODE_PROVIDER=openai. Ollama needs no key.
`OPENROUTER_API_KEY`	Optional: OpenRouter key; used instead of a TypeSafe key.
`AI_GATEWAY_API_KEY`	Optional: Vercel AI Gateway key; used instead of a TypeSafe key.
`JEV_CODE_PROVIDER`	Optional: Forces a host: typesafe, openrouter, vercel, openai or ollama (needed for the last two).
`OPENAI_API_KEY`	Optional: OpenAI Decisions API key; used only when JEV_CODE_PROVIDER=openai.
`OLLAMA_HOST`	Optional: Ollama server address; localhost:11434 by default.
`TYPESAFE_BASE_URL`	Optional: A proxy or your own System One gateway; the client appends /v1/systemone.
`TYPESAFE_DEFAULT_MODEL`	Optional: Pins a model, for example jev-1.13; the default depends on the host.
`JEV_CODE_TIMEOUT_MS`	Optional: Timeout per attempt in ms; 30000 by default.
`JEV_CODE_MAX_RETRIES`	Optional: Retries on 429, 5xx, timeouts and connection errors; 2 by default.

What it does

jev-code is a stdio MCP server, plus a CLI and a Pi extension, that gives a coding agent five tools backed by Jev, TypeSafe's decision model. The agent sends the evidence (test failures, files, findings, a PR description) along with its own labels, questions or scale. What comes back is structured JSON: a label or verdict per item, the probability distribution, and an `auto` or `review` decision worked out from thresholds the caller can change on each call. Jev returns no free text.

It suits the small sorting jobs an agent does during coding work, such as splitting CI failures into flaky and real or ranking which files answer a question. Each call puts all its items into one request. Input is checked locally first (shapes, duplicate ids, request size), so a malformed payload is rejected before it costs a request. The package also includes an Agent Skill that tells the agent when to use a classifier, and a `setup` command that registers the skill and the server in Claude Code, Codex, Pi and OpenCode.

The model runs on a remote host, not on your machine. Supported hosts are TypeSafe, OpenRouter, Vercel AI Gateway, OpenAI's Decisions API (opt-in, public beta) and a self-run System One gateway. With Ollama 0.35 or later it can use local decision models and needs no key. Its SECURITY.md documents `OLLAMA_API_KEY`, which the README leaves out: it is sent only to an Ollama server that requires one. The same file says that only the latest published minor version gets security fixes.

What it needs

Runtime	Node.js 20+

Node.js 20 or later, run through npx. You need one model host: a TypeSafe, OpenRouter or Vercel AI Gateway API key; an OpenAI key plus JEV_CODE_PROVIDER=openai; or a local Ollama 0.35+ server with a decision model pulled (for example `nimble`) plus JEV_CODE_PROVIDER=ollama. Every remote host bills calls to your own account.

Limits

The tools are text-only on every host, even where the model accepts images. Cloudflare Workers AI serves Jev but is not supported. On OpenAI's Decisions API (public beta) the model can refuse a single item, which then comes back as `review` or `uncertain`, and retrying the same input does not change that. OpenAI is never used just because OPENAI_API_KEY is set; you have to choose it. Requests time out after 30 s per attempt by default and are retried twice; JEV_CODE_TIMEOUT_MS and JEV_CODE_MAX_RETRIES change this. The project is not affiliated with TypeSafe. Installs made with 0.2.1 or earlier run an unpinned command and do not pick up updates until setup is run again.

Who it suits

Developers using Claude Code, Codex, OpenCode or Pi who want calibrated labels and probabilities for many items at once, rather than spending a general-model call on each item. It does not read your files or repo by itself: it only classifies what the agent passes in. Your evidence goes to the remote host you pick, so for code that must stay on your machine, use the Ollama mode. Its README names jkudish/jev-mcp (held here as jev-mcp-2) and itsmostafa/typesafe-mcp as other Jev MCP servers with different tool sets. Choose jev-code if you also want the bundled skill, the one-command setup and the choice of hosts.

Tools 5

jev_classify: Gives each item one label from classes the caller defines, with probabilities, margin and an auto/review decision.
jev_check: Answers yes/no questions about one piece of evidence with a probability and a yes, no or uncertain verdict.
jev_score: Places many items on one ordered scale, such as severity or priority, with score, nearest level and confidence.
jev_rank: Scores the relevance of each candidate to a question, returns them sorted, and says whether any is relevant.
jev_ask: Sends mixed question types about one state and returns the raw System One answers.

Popularity

44 popularity score, on our list (the bar is 36)

★ 58 ↓ 702/wk

GitHub stars	58	21.3 points
npm downloads, last week	702	22.8 points

Measured 2026-11-07. How the score works.

Where it is published

Code	francoischastel/jev-code ★ 58
npm	@french-castle/jev-code 0.5.0, 2026-10-08

What it can reach

Runs	On your machine
Reaches	The internet
Changes things	No: it reads only
Sign-in	An API key or token
Licence	MIT (npm)
Last code change	2026-10-08

The tools run on your machine and send one POST to /v1/systemone per call. The request goes to the host your key belongs to (api.typesafe.ai, openrouter.ai, ai-gateway.vercel.sh, or api.openai.com when chosen), or to localhost:11434 for Ollama. It carries only the payload the agent passed in; the server does not read files, git state or session history by itself. A key is sent only to the host that issued it, and a base URL pointing at a different known host is refused. The server's stderr line shows keys masked. The tools change nothing. The separate `setup` command does write: it creates skill folders and harness config entries (backing up the existing files first), runs the claude, codex and pi CLIs, and copies the key in use into those configs unless --no-env is given.

Over time

GitHub stars 58 +14 (+32 %)

5844

13 Oct 20267 Nov 2026

npm downloads a week 702 −155 (−18 %)

1.0k224

5 Oct 20267 Nov 2026

Before the site measured it, read from npm's daily download counts.

Popularity score 44 +1 (+1 %)

4521

13 Oct 20267 Nov 2026

What is rising across the directory

Sources

Page checked 2026-10-14.

Badge for your README

Shows this server's rank in its category, and stars for a server that runs on your machine. It updates every hour.

MarkdownREADME.md

[![jev-code on EveryMCPServer](https://everymcpserver.com/badge/jev-code.svg)](https://everymcpserver.com/servers/jev-code)

Connect your client

Add in one click

Or add it by command or config

Settings 10

What it does

What it needs

Limits

Who it suits

Tools 5

Similar servers

Jev MCP

LLM Checker

Ruflo

ruv-swarm

Agentic Flow

ReevesAgents