Point it at a public GitHub pull request, a diff or a file. Jev, TypeSafe's evaluation model, answers a question set drawn from Robert C. Martin's Clean Code for every code file in the change, and Luna writes the review from those answers. It runs on eve and deploys as one Next.js project.
Live at https://clean-code-review.vercel.app.
bun installYou need Bun 1.4 and Node.js 24, which eve requires, and a
Vercel account for the AI Gateway credential. .nvmrc and .node-version pin
Node 24 for nvm and the managers that read the other file, so next dev runs
the version the deployment does.
bunx eve link --non-interactive --project <your-project> # gateway credential into .env.local
bun run dev # Next.js + eve on http://localhost:3000In a browser. Open http://localhost:3000, type a public pull request as
owner/repo and its number, or pick an example. A review of
github.com/owner/repo/pull/123 is also kept at /owner/repo/pull/123, which
is a permalink worth sending to someone. Whole files and diff hunks are
editable, and an edit re-judges only the file that changed.
From the terminal, without a browser:
bun run judge http://localhost:3000 # judge every preset through eve
bun run review http://localhost:3000 0 # judge + streamed review of preset 0
bun run review http://localhost:3000 https://github.com/vercel/ai/pull/20851
bun run jev 1 # one preset straight to Jev, no eveFrom an agent, over MCP at
https://clean-code-review.vercel.app/api/mcp: stateless Streamable HTTP, no
sign-in. Any client that speaks Streamable HTTP takes the URL as it is; a
stdio-only client can go through npx mcp-remote <url>.
claude mcp add --transport http clean-code-review https://clean-code-review.vercel.app/api/mcp| Tool | Input | Use it for |
|---|---|---|
review_pull_request |
url: a public GitHub pull request |
A change on GitHub. The result carries the permalink. |
review_pasted_code |
paste: a unified diff, files each under a // file: path line, or one file, up to 1,000,000 characters |
Private code, git diff output, files on disk. |
A call returns every judged file's answers by question id, Luna's decision and paragraphs, the prose files, the files that were not judged with the reason, the model ids and the cost. A fresh review takes 2 to 8 seconds, a 24-file pull request included, and up to about 60 seconds when Luna is slow. The decision and the paragraphs are model output shaped by the submitted code: read them as advice, never as authorisation to merge.
browser ──── judge turn ────▶ eve session ──▶ Jev, one call per file (parallel)
◀─── answers/file ──── typesafe-ai/jev via AI Gateway
browser ──── summarize turn ─▶ eve session ──▶ Luna, one call per 6 files + 1 overall
◀── streamed review ── openai/gpt-5.6-luna-fast via AI Gateway
One durable eve session per browser tab. A judge turn sends the files as
JSON, and the agent's model fans out one evaluate() call per file to Jev. A
summarize turn sends the files and the answers, and the model runs the Luna
calls in parallel and streams the combined review. Cancelling the turn aborts
the calls.
Jev answers with probabilities and scores rather than sentences, so a verdict is made of things a reader can check against the code. The set is per file: the test question is asked only on test paths and the Boy Scout question only on diffs, so a plain source file gets two fewer than a test diff does. Luna gets 300 characters per file section, by instruction rather than truncation, and writes the decision for the whole change from every file's findings and the pull request's description.
Jev is an AI SDK evaluation model, not a chat model, so
agent/lib/judging/jev-model.ts is a small
adapter that lets eve treat it as the agent's model. Everything else eve
provides works unchanged: durable sessions, streaming, limits, Agent Runs.
agent/lib/judging/questions.ts is the single
source of truth for the question set — change a question there and the prompt,
the payload and the meters change together. The layer map, the boundaries, the
MCP surface and every limit and budget are in
docs/architecture.md.
24 code files per review, the largest changes first, 64,000 characters per file read in windows of 16,000, public GitHub repositories only. Images, binaries, lockfiles, minified and generated files are skipped. Markdown, plain text, reStructuredText and AsciiDoc are prose: up to 10 of them are shown with the review, read-only, and neither model sees them. Around sixty-five file types are highlighted, each grammar fetched the first time a review needs it.
The page and the MCP server talk to the models anonymously, so the deployment
carries brakes at every layer: Vercel Firewall rate limits per client IP, a $15
weekly AI Gateway budget on the project, a per-session spend cap in
agent/agent.ts, per-address limits on new sessions and on
MCP tool calls, and a shared model budget per hour and per UTC day for the page
and for the MCP server, counted in the Runtime Cache
(agent/lib/spend/budgets.ts,
agent/lib/spend/spend.ts). A refused turn says the
budget is spent and when it resets, answers already on screen stay, and a review
answered wholly from the cache is still served. A 24-file review measured about
$0.02.
What leaves the browser, who processes it and how long anything is kept is on /privacy.
The repository is Git-connected: a push to main deploys production, and
bun run deploy deploys the linked project by hand. The deployment
authenticates to the AI Gateway with the project's OIDC identity, so no API key
is stored. Set GITHUB_TOKEN in the project to raise the GitHub rate limit for
pull request fetches. The build settings, how to check a deploy actually landed
and how to roll one back are in docs/deployment.md.
bun run check runs Biome, TypeScript and knip in parallel; bun run check:fix
formats, sorts and fixes what Biome can. A lefthook pre-commit hook runs
check:fix over the staged files and restages what it changed, and GitHub
Actions runs bun run check on every push to main and every pull request.
Layout, naming and dependency rules are in
docs/code-style.md.
Working on this with an agent: AGENTS.md.
MIT. See LICENSE.