@maximem/jev-mcp

MCP server exposing TypeSafe Jev as typed, calibrated decision tools for coding agents — plus a harness that gates the agent's own work.

Jev doesn't generate text. It answers a map of typed questions about a state and returns calibrated probabilities in about 100ms.

Install

Claude Cowork — add to claude_desktop_config.json (Claude Desktop passes stdio servers through into Cowork's sandbox), then fully quit and reopen Claude Desktop:

{
  "mcpServers": {
    "jev": {
      "command": "npx",
      "args": ["-y", "@maximem/jev-mcp"],
      "env": { "TYPESAFE_API_KEY": "ts_live_..." }
    }
  }
}

On Windows, wrap npx in cmd /c.

Claude Code:

claude mcp add jev --env TYPESAFE_API_KEY=ts_live_... -- npx -y @maximem/jev-mcp

Get a key at console.typesafe.ai. It is read from the environment only — never passed as a tool argument.

Tools

Primitives: jev_classify, jev_score, jev_check, jev_rank, jev_ask, jev_health

Harness: jev_plan_gate, jev_action_gate, jev_verify, jev_screen, jev_select

Every response is MCP structured content carrying the full probability distribution, a confidence, the resolved model version and measured latency.

When not to use Jev

• Checkable by code — file exists, tests pass, it compiles → run the command, not a model.
• Open-ended writing or synthesis → your frontier model.
• Trivial work → skip the harness. Gating a one-line edit costs more than it saves.

JEV_HARNESS=off disables the automatic checkpoints; the tools stay callable.

Full documentation, configuration and the decision table: github.com/maximem-ai/jev_mcp

MIT