Jev MCP — npmjs.com

Run npx -y @jkudish/jev-mcp, or register it with your client, for example claude mcp add jev -- npx -y @jkudish/jev-mcp. Node.js 22 or newer is required. The README also gives config snippets for Amp, Codex, OpenCode, and any other MCP client.

It needs TYPESAFE_API_KEY for a Jev provider; TypeSafe direct is the default and alternatives are listed under Configuration. Some MCP clients filter the environment before spawning servers and silently drop the key, so pass it explicitly in the server config if the server reports a missing key.

Twelve: jev_verify, jev_screen, jev_noul, jev_find, jev_rerank, jev_classify, jev_decide, jev_compare, jev_extract, jev_audit, jev_review, and jev_gate. They cover claim checking, pre-context screening, calibrated probabilities, candidate selection, reranking, batch classification, bounded decisions, passage comparison, field extraction, and patch review.

Yes. Run it with --http (or JEV_MCP_TRANSPORT=http) to serve stateless MCP at /mcp with a health check at /health. PORT defaults to 8080 and HOST defaults to 127.0.0.1; set HOST=0.0.0.0 to serve beyond your machine. JEV_MCP_AUTH_TOKEN is required unless HOST is loopback.

The README states each judgment comes back in roughly 150 to 500 ms for a fraction of a cent, which is why the mechanical checks are cheap enough to run on every page, claim, or candidate list. Every call spends your own Jev key.

It is MIT licensed and published on npm as @jkudish/jev-mcp by jkudish. The README describes it as early software with rough edges and invites issues and pull requests, pointing to CONTRIBUTING.md.

It is early software with rough edges. Clients that cannot send headers, such as claude.ai custom connectors, need JEV_MCP_PATH_TOKEN=1 and a URL-safe token in the path, served over HTTPS only, with the whole URL treated as a secret and rotated if it escapes.