Jev MCP Server已精选

What is Jev MCP?

Jev MCP is a community-built Model Context Protocol server that gives compatible agents access to TypeSafe AI's Jev evaluation model. Jev is designed for bounded judgments rather than open-ended text generation: an application supplies state plus typed questions, and receives structured answers with probability distributions and confidence information.

This server packages that decision layer as MCP tools for claim checking, prompt-injection screening, semantic ranking, and batch classification. It can be connected to Codex, Claude Code, Cursor, OpenCode, and other clients that support local stdio MCP servers.

Jev MCP is maintained by jkudish as an independent open-source project. It is not an official TypeSafe AI MCP server. The underlying Jev model and direct API are provided by TypeSafe AI.

For a model-level explanation, see What is Jev?.

Available tools

jev_verify

Checks one or more claims against supplied evidence. It returns a verdict for each claim, probability information, confidence, and a recommendation to accept automatically or send the result for review. It is useful for reports, PR descriptions, research briefs, and citation-aware workflows.

jev_screen

Evaluates text before an agent places it in context. The response estimates prompt-injection risk, substance, and relevance, then recommends pass, review, block, or skip. The recommendation is advisory: the calling application still owns enforcement and permission decisions.

jev_find

Ranks a bounded set of candidates against a plain-language query without requiring a persistent embedding index. It also checks whether the candidate set contains an answer at all, which helps avoid treating the least-bad result as a valid match.

jev_classify

Assigns multiple items to a shared set of classes in one request. Each item receives a class, probabilities, confidence, and a decision signal based on configurable acceptance and margin thresholds.

Install Jev MCP

Jev MCP requires Node.js 20 or newer. The package can be launched directly from npm:

npx -y @jkudish/jev-mcp

Claude Code users can register it with:

claude mcp add jev -- npx -y @jkudish/jev-mcp

For Codex, add this to ~/.codex/config.toml:

[mcp_servers.jev]
command = "npx"
args = ["-y", "@jkudish/jev-mcp"]

Other MCP clients can use the JSON configuration shown in the installation section above. Set the API key in the MCP server environment instead of placing a real credential in a prompt or checked-in configuration file.

Authentication and providers

The default direct provider uses TYPESAFE_API_KEY, available from the TypeSafe console. The project also documents provider selection through OpenRouter, Cloudflare Workers AI, and Vercel AI Gateway. Provider behavior, model aliases, pricing, and environment variables can change quickly, so check the repository before deploying a pinned production configuration.

Safety and limitations

Jev returns typed judgments, not guaranteed truth. A valid schema means the result is structurally usable; it does not make the classification or probability infallible. Test thresholds on representative data, keep authorization and irreversible side effects in ordinary application code, and route uncertain or high-impact decisions to a person or a stronger review path.

The screening tool is also not a security boundary by itself. A calling agent or harness must enforce the returned recommendation and must continue to apply normal sandboxing, least privilege, URL validation, and secret-handling controls.