Jev MCP server: a second opinion, with probabilities, for any agent

The Jev MCP tools an agent gets

Three general judgments and a comparison cover almost anything an agent wants checked. The other three tools on the server grade essays and resumes and are described on the grading API page.

Each result arrives twice: a one-line summary the agent's model can read at a glance, and the full numbers as JSON for code that wants a threshold of its own.

When an agent should ask Jev: five guard patterns

The Jev MCP server earns its keep at the moments an agent is about to trust itself. These are the five we see most, each with the arguments an agent would pass.

1. Before an irreversible action. Ask judge_yes_no whether the planned command can be undone without losing data, with the command and the directory listing as the text. Below 0.6, stop and ask the person.
2. After a tool call. A command can exit cleanly and still fail. Pass its output and ask whether it shows success, so the agent does not report a win the log contradicts.
3. Before claiming the work is done. Give the test output and ask whether every test passed. An agent's own summary is the least reliable source for that answer.
4. Choosing a tool. judge_choice with each tool as an option and a line on when it applies; route on the winner when the confidence is high, and let the model reason when it is not.
5. Before the reply goes out. judge_score with the user's question and the draft, on levels from “does not answer” to “fully answers”. Below the top level, revise once.

{
  "tool": "judge_yes_no",
  "arguments": {
    "text": "Planned: rm -rf ./build ./data\nls: build/ data/ notes.md",
    "question": "Can this action be undone without losing data?",
    "yes_means": "Only generated files are removed",
    "no_means": "Source files or data would be lost"
  }
}

Every one of these is a bounded question with a right answer, which is what Jev is built for, and each costs a credit or two — cheap enough to run on every step rather than only the ones that already look risky.

Connecting the Jev MCP server

Add https://jev-agent.com/api/mcp to your client as a remote HTTP server and send your key as Authorization: Bearer jv_live_…. The exact lines for Claude Code, Cursor and VS Code, and the mcp-remote bridge for clients that only run local servers, are on the grading API page. A key comes from API access; the same key works for the REST API.

Other Jev MCP servers

Four community servers call Jev with a key of your own and run on your machine. The figures are from GitHub on 2026-10-02.

jkudish/jev-mcp is the one most people use and the most actively kept: it was updated on the day we checked and ships the widest set of tools. Jevbridge is the unusual one, because it can run without a TypeSafe key and also speaks ACP for editors such as Zed. The other projects in the ecosystem are on the open source page.

Hosted or self-run: which Jev MCP server to pick

Pick the hosted Jev MCP server when you have no TypeSafe key yet — keys come from a waitlist — or when you want one key and one balance across the MCP server, the REST API and the graders. Pick a self-run server when you already hold a TypeSafe key, want the call to go straight from your machine to TypeSafe, or need a tool the hosted one does not have, such as reviewing a diff. Nothing stops you running both: they are separate servers with separate names in your client.

Jev MCP questions