JEV GUIDE

Jev MCP setup: connect a decision tool to your coding agent

MCP can expose Jev judgments to a coding assistant. This guide uses a reviewed typesafe-mcp revision and separates transport setup from successful model execution.

Jev Gallery · Reviewed 2026-09-19

Choose the tool you actually need

typesafe-mcp offers a general evaluate tool taking state and questions. The jev-mcp case covers citation checks, content screening and candidate reranking. Similar names do not imply identical configuration; inspect the linked cases.

For example, an assistant could check whether feedback describes a concrete defect before the application opens a handling step. The MCP server is not itself a complete action-execution system.

Source: typesafe-mcp · source at reviewed revision ↗

Verify three separate layers

First confirm that the executable starts. Then verify MCP initialization and tool discovery. Only a real evaluate request with a valid credential tests model access.

Review platform support, installation instructions and a fixed version in an isolated directory.
Configure the command and environment in your chosen MCP client.
Confirm evaluate appears in the tool list, then try a synthetic input.
Record the actual response, error and usage before connecting a real workflow.

What the tool arguments look like

This is a tool-argument example, not a complete JSON-RPC message. It checks a synthetic release note for a migration requirement. It reads no local repository and includes no live model result.

JSON

{
  "name": "evaluate",
  "arguments": {
    "state": {
      "release_note": "Version 2 changes the settings format. Run the migration command before restarting."
    },
    "questions": {
      "migration_required": {
        "type": "noul",
        "instructions": "Does `release_note` explicitly require a migration before using the new version?"
      }
    }
  }
}

Source: typesafe-mcp · source at reviewed revision ↗

Avoid configuration confusion

This revision supports TypeSafe and OpenRouter routing. Check the actual endpoint, model and billing provider, particularly when another provider’s variable is already present.

The setup command changes client configuration. Back up existing settings and check names before using it. Supply credentials through the local environment or server configuration, and keep them out of shared screenshots.

Source: typesafe-mcp · source at reviewed revision ↗

What has been checked here

The pinned revision passed local Go tests, build, MCP initialization and tool discovery. No real evaluate call was made. Those checks do not establish model quality, network latency or cost; live access still requires a valid account credential.