jev-mcp

What it does

Jev MCP, by Joey Kudish (who also makes jev-browser), is an MCP server that turns common Jev patterns into ten ready-made tools:

Tool	What it does
jev_verify	Checks claims against evidence, one verdict per claim
jev_screen	Judges fetched text for prompt injection, substance, and relevance before the agent reads it
jev_find / jev_rerank	Picks or ranks candidates by meaning, and reports whether any candidate answers the query at all
jev_classify	Sorts up to 64 items into your own classes
jev_decide	Chooses between 2 to 6 alternatives, with requirement checks and an “ask the user” escape hatch
jev_compare	Judges how two passages relate, aspect by aspect
jev_extract	Finds candidate values with your regex, then has Jev pick the right one, returned verbatim
jev_review / jev_gate	Scores a proposed diff, and checks completion claims such as “tests pass” against the evidence you supply

Each tool builds the right Jev questions for you, validates the answers, and returns a compact result with a recommended action, such as auto versus review, or pass, review, block, or skip. Most thresholds are parameters. Many tools follow patterns from TypeSafe’s cookbooks and link to them.

Setup

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

It needs Node 20 or newer and TYPESAFE_API_KEY. There are setup snippets for Amp, Codex, OpenCode, and any MCP client, and a paste-in prompt that asks your agent to install it without putting the key in the chat. It can also route Jev calls through OpenRouter, Cloudflare Workers AI, or Vercel AI Gateway. With a TypeSafe key set, TypeSafe is used first.

Using it

• The checks pass. Type checking and all 53 tests passed. CI runs them on every push to main and on pull requests.
• All ten tools worked. We called each one over stdio with small, realistic inputs and a real Jev key. Every call returned a structured result with the model used, the provider, token usage, and a recommended action where the tool has one. None errored.
• Extraction stays verbatim. jev_extract pulled a version number from release notes exactly as written. Jev only chooses among regex matches, so it can’t invent a value.
• A hostile regex can’t hang it. We gave jev_extract a pattern with catastrophic backtracking. It ran in a separate worker, was stopped after one second, and came back as invalid_pattern with a clear message, while the rest of the call and the following calls worked normally.
• Input limits are explicit. Tools cap item counts, candidate lengths, and document size (50,000 characters for extraction), and say so in their schemas, so an agent learns the limits before it hits them.

As with all our reviews, we don’t publish Jev’s answers or confidence values from our test calls. The README’s own examples and timings are the author’s.

Permissions and data

• It sends what your agent sends. There’s no file or network access beyond the Jev API. Each tool forwards the text it’s given, whether that’s claims, pages, documents, diffs, or candidates, so your agent decides what leaves your machine.
• It advises, it doesn’t enforce. jev_screen might recommend block for a page with injected instructions, but the server never blocks anything itself. Your agent has to follow the recommendation.
• The provider is picked automatically. If you don’t have a TypeSafe key but do have a Cloudflare token and account ID in the environment, for example from other tools, it will route Jev calls through Cloudflare. Set JEV_PROVIDER to be explicit.
• The tools don’t declare MCP hints. None of them marks itself read-only, so clients can’t tell from the tool definitions that nothing is modified.

Maintenance

Twenty-six commits and five releases between September 17 and 19, with a changelog, CI, a security policy, and a contribution guide. It’s MIT licensed. The review and gate tools came from a community proposal that the author reworked and merged. No issues are open.

Verdict

Jev MCP is the broadest Jev toolkit for agents we’ve tested: ten tools where the other MCP servers we’ve reviewed offer one. The tools are focused, their limits are explicit, and the recommendations are advisory, so your agent stays in control. It’s a good default if you want your agent to fact-check, screen, rank, or extract with Jev without writing questions yourself. Keep in mind that everything passed to a tool goes to TypeSafe. For a single general-purpose evaluate tool that you shape yourself, see System One Connector. For code-quality scores specifically, compare jev_review with Jev Review MCP.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 19, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.