Install it, start it
Two commands, and the setup opens in your browser. It checks that Claude or ChatGPT is installed and signed in. One AI is enough; two cross-check each other.
Self-hosted AI code review
Lookover runs on your own computer, reviews your pull requests with Claude or ChatGPT through their official CLIs, and posts what it finds on GitHub. Built for coding agents: they wait for the review, fix, and ask again.
No account to create. Needs a Claude or ChatGPT plan, git, gh and Node.js 18 or later. Windows (x64), macOS, Linux.
a1f9c2e · round 1One thing to fix. The rest can go.
Found by Codex, double-checked by Claude.
API key is written to the request log. When a client is throttled, the full headers are logged, including Authorization. Log the client id instead.
b2c3d4e: logs the key id only.Resolved · verified fixed in b2c3d4eb2c3d4e · round 2One pull request, two rounds. A condensed preview of the review, the fix and the passing check on GitHub.
Reviews use the Claude or ChatGPT plan you already have. No per-review price, no credits to buy, no seats.
On your computer. No account with us, no hosted service in between, no telemetry.
It runs the official claude and codex CLIs, signed in as you. It never reads or stores your tokens.
Setup
Three steps. The first review posts for real, on a pull request you pick.
Two commands, and the setup opens in your browser. It checks that Claude or ChatGPT is installed and signed in. One AI is enough; two cross-check each other.
Create your App: a GitHub App on your account or an organisation's, which posts the reviews as itself. Then you pick the repositories it reviews.
Each one gets a review on GitHub. Lookover keeps a list of where each stands; fix, push, and ask for another look, or let your agent ask.
How it works · built for coding agents
Every review ends with a check on the commit and findings an agent can read. One command, lookover wait, does the waiting and says what to do next. Here is one pull request, from the agent's terminal to the merge button, with Lookover's own screen beside it.
No open pull requests.
Looked over a1f9c2e. One thing to fix, rateLimit.ts line 42. The rest can go.
Looked over b2c3d4e. Nothing new to fix. The rest can go.
src/middleware/rateLimit.ts line 42When a client is throttled, the full headers are logged, including Authorization.
b2c3d4eResolved · verified fixed in b2c3d4e ✓The loop settles. "Fixed" is checked on the new commit, never trusted. Nothing runs on half-finished pushes: the agent asks when it is ready. When the agent and the reviewer cannot agree, the check says a person is needed, and the agent stops.
Works with the agents you run. A skill makes the loop one sentence for Claude Code or Codex: "get this pull request through review". Anything that can call gh can do the same by hand. Setting up your agent, in the docs.
What it posts
A reviewer that posts twelve nits gets muted. Lookover is built to post the two that matter.
Code written by Claude is looked over by Codex, and the reverse. Models miss their own habits; a second model catches them.
Unless you pick Quick, every finding is re-checked in a fresh session before it reaches you: one ruled out is dropped, one left in doubt is marked unconfirmed. Posted: 2 of 5.
On the next push it checks each earlier finding: fixed, still there, or no longer relevant. It only posts what is new. An empty review is one line.
Five minutes to set up. No account to create.
Questions
Nothing beyond the Claude or ChatGPT plan you have. A review draws on your plan's limits the way a session in the terminal would. Lookover reviews one pull request at a time per AI, uses no API key, and waits for your plan's limit to reset rather than paying past it. If your plan can buy extra usage, leave that off.
Claude Pro or Max, through the claude CLI; ChatGPT Plus or Pro, through the codex CLI. One is enough. With both, the AI that did not write the code looks first and the other double-checks.
Only where it already goes: to GitHub, and to Claude or ChatGPT through their own CLIs. Nothing to us. No account, no telemetry; the source is open, so you can check. Where your code goes, in the docs.
One review in a few words with the findings on the lines they are about, one summary comment that stays up to date, and a check on the commit. An empty review is one line. It never merges, approves or requests changes.
Lookover only runs the official claude and codex CLIs, signed in through their own flows, on your machine, for your own pull requests. It never reads or stores your tokens. Your plan's terms apply: one subscription, one person, so each colleague runs their own. Lookover is not affiliated with Anthropic or OpenAI. Your plan's terms, in the docs.