jev-mcp

MCP server that exposes TypeSafe AI's Jev decision model to Claude (Desktop, Code, Cowork). Two tools:

Tool	Use
jev_evaluate	Full System One call: state + any mix of noul / choice / score questions
jev_route	One choice question + confidence threshold → {choice, confidence, escalate}

Setup

Needs Python 3.11+ (macOS ships 3.9, so name the version explicitly).

python3.12 -m venv .venv                  # or: uv venv --python 3.12
.venv/bin/pip install -e ".[dev]"         # or: uv pip install -e ".[dev]"
cp .env.example .env                      # add your JEV_AI_API_KEY
chmod 600 .env                            # it holds a secret
.venv/bin/pytest                          # mocked, no network needed

The server reads JEV_AI_API_KEY from the environment, falling back to this .env. Keeping the key here means it lives in one file instead of being copied into every MCP client's config — so the registrations below carry no secrets.

Try it

npx @modelcontextprotocol/inspector jev-mcp

Add to Claude Desktop

claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "jev": {
      "command": "/absolute/path/to/jev-mcp/.venv/bin/jev-mcp",
      "args": []
    }
  }
}

Fully quit Claude Desktop (⌘Q) and reopen — the config is only read at launch. Add an "env": { "JEV_AI_API_KEY": "..." } block only if you want to override the key in .env; environment variables take precedence.

Add to Claude Code

claude mcp add jev -s user -- /absolute/path/to/jev-mcp/.venv/bin/jev-mcp

-s user makes it available in every project; the default (local) registers it only for the current directory. Restart the session afterwards — MCP servers are loaded at startup.

API reference

Docs: https://docs.typesafe.ai/api — POST /v1/systemone, verified 2026-09-21. Live API base: https://jev-ai.pro/api (override with JEV_AI_API_BASE). Note the docs host and the API host differ; api.typesafe.ai rejects sk_jev-ai_ keys. Jev 1.13 limits: 64k total context, 32k for state + longest question.