Jev In MCP

It is an MCP relay: to the agent it is one MCP server, and to your MCP servers it is a client. Every tool passes through unchanged, every server gets one extra <server>__use_jev tool, and there is one global use_jev over all servers, each handing a goal to TypeSafe Jev.

It calls <server>__use_jev(goal, max_steps?, pause_after?, session?, input?) to run that server's tools toward the goal, or use_jev(goal, servers?) to work across every server, where Jev picks the server first and then the tool in one request.

With Node 20+, install from GitHub while it is not on npm, run jev-in-mcp setup to store the key locally, jev-in-mcp import to copy mcpServers entries from Claude Code, Cursor, Windsurf and Claude Desktop, and jev-in-mcp status to see what is configured. Then replace your client's servers with one jev entry.

Configuration is at ~/.config/jev-in-mcp/config.json, or $JEV_IN_MCP_HOME if set, and the Jev API key sits in credentials.json next to it with mode 600, or in the JEV_API_KEY environment variable. Setup stores the key through a local page, not a chat.

The call returns status needs_input with the field and a session, and the calling model calls the same tool again with session plus input so the run continues. Where the client supports MCP sampling, the relay asks the model inside the call instead and nothing is interrupted.

Version 0.1.1 was verified live against the reference filesystem server with real Jev and DeepSeek as the calling model: 6/6 tasks in both the needs_input and sampling paths, 4 to 10 seconds per task, and the same 6/6 through Vercel AI Gateway. Other servers have not been tried yet.

Tools with a required parameter Jev cannot express, such as objects or arrays, are not offered to it. The deny option keeps destructive tools away and describe replaces a tool's sentence with what actually happens, which the README says is where the quality of a server's use_jev is decided.