Jev Review MCP

Not to be confused with Jev Review by Dev Agrawal, a separate code-review pipeline with a local dashboard.

What it does

Jev Review is a local MCP server with one tool, jev_review, for Claude Code, Codex, Cursor, and OpenCode. Your agent sends the task, a focused diff, and optionally whole files and repository notes. Jev Review then asks Jev 57 questions in a single request, three for each of 19 quality dimensions: whether the dimension can be judged from the context, a 1–10 score, and the most important weakness from a fixed list. Dimensions include correctness, readability, security, test quality, and, when relevant, performance and compatibility.

The result has a score and confidence for each dimension that applies, applicable: false for the rest, a priority list, and short hints taken from predefined text. There’s deliberately no overall grade and no prose explanation. On a follow-up call, the agent passes the previous result back, and the tool reports improvements, regressions, and a comparison per dimension. The bundled skill turns this into a loop: implement, score, investigate, improve, and rescore.

Setup

export JEV_API_KEY="your-key"
npx plugins add NiazMorshed2007/jev-review

It needs Node 20 or newer. There are plugin manifests for Claude Code, Codex, and Cursor, and manual instructions for OpenCode. It isn’t published to npm. Clients run the committed dist/server.js bundle directly from the repository.

Note the variable name: it reads JEV_API_KEY, while TypeSafe’s SDKs and most other Jev tools use TYPESAFE_API_KEY.

Using it

We drove the server over stdio with a small scripted MCP client, the way a coding agent would, using a real Jev key.

• The bundle matches the source. Rebuilding dist/server.js from source produced a byte-for-byte identical file, so what clients run is what you can read. Type checking, the 13 tests, and the build all passed.
• The loop works as described. A first call on a small change returned all 19 dimensions. For a change that small, only a handful were scored, and the rest were marked not applicable rather than given a made-up number. After we changed the code and passed the first result back, the response added improvements, regressions, and a per-dimension comparison.
• Errors are clear. An empty call, a missing key, and an invalid key each produced a plain message telling the agent what to fix.
• It’s honest about its limits. The README says the scores aren’t root-cause explanations, and the skill warns against gaming them with needless abstraction, tests, or comments.

As with all our reviews, we don’t publish Jev’s scores for our test code.

Permissions and data

• It reads nothing on its own. Only what the agent passes (task, diff, files, and notes) goes to api.typesafe.ai. The tool is marked read-only and never edits files. There’s no other backend or telemetry.
• But it’s designed to be called often. The server’s instructions and the skill tell the agent to use it “for every nontrivial coding task”: once for a baseline and again after each improvement. Installing it means your diffs, and sometimes whole files, go to TypeSafe during normal work, not only when you ask for a review.
• Secrets are kept out only by instruction. The skill tells the agent never to send secrets or environment files, but nothing in the code filters them.
• The model can’t be pinned yet. Requests always use jev-latest, so scores may shift when TypeSafe updates the model. An open pull request from a contributor makes the model configurable.

Maintenance

Four commits on September 17, MIT licensed, with a small test suite but no CI. There’s one open feature proposal and the model-configuration pull request, neither answered yet.

Verdict

Jev Review is a clean, well-scoped way to give a coding agent a repeatable quality signal, and its per-dimension comparisons are more useful than a single grade. It works best when you want your agent to self-check and iterate, and you’re comfortable sending your working diffs to TypeSafe throughout a session. If you want review only on demand, tell your agent explicitly when to call it: the server’s own instructions ask for a review on every nontrivial task, so removing the skill alone isn’t enough. For a pipeline that screens a finished diff and suggests who should look at it, see Dev Agrawal’s Jev Review. For a general Jev tool your agent can call, see System One Connector.

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.