Agent-written pull requests wait 5.3 times longer for a reviewer.
Agents open pull requests faster than any team can read them.
Source: LinearB, 2026 benchmarks
The Quality Engine for AI generated code
Cerebe reviews pull requests and posts the result as a GitHub check. Models from competing vendors review each change separately.
See how it works →Setup instructions
Paste this prompt into your coding agent. Full installation guide →
Install Cerebe in this repository. Follow https://cerebe.ai/setup.md in order. Lead with the GitHub App at https://github.com/apps/cerebe-factory/installations/new, then the CLI, and finish by running cerebe login. When cerebe is already on PATH, run cerebe --help. If login is not listed, run the installer before cerebe login. When login is listed and this checkout has no cerebe/ directory, run cerebe install instead of the installer. When this checkout already has a cerebe/ directory, run cerebe init.
- Check
- cerebe/critic on src/auth.ts, line 29: res.cookie("session", session.token, { secure: true, sameSite: "lax" })
- Blocker, contracts
- A successful refresh invalidates the caller's bearer token and never returns a replacement the API will accept. Required fix: return the new token in the JSON body with the same contract as login.
- Medium, security
- The cookie stores the rotated session token without httpOnly, so any script on the page can read it. Required fix: if a cookie is kept, set httpOnly and a maxAge from the session TTL.
Each finding explains why the issue matters and how to fix it. When an issue is in a specific file, the finding names the file, and the line when there is one.
On your workstation and GitHub
The CLI reviews every commit on your workstation alongside Claude Code, Codex
and Cursor, catching problems before you open a pull request.
It works with whichever AI coding harness, model and cloud you already use.
On GitHub, the cerebe/critic check is the approval of record.
Models identify issues. Fixed rules determine the verdict from their findings. Each model reviews without seeing the others' findings. The same rules apply locally and on GitHub, and each commit gets one verdict. Require cerebe/critic in the repository's ruleset to gate merges on its verdict.
Reviews use your AGENTS.md and Cerebe's rules. Customer-owned rules and doctrine are on the roadmap. Review agents run in an isolated sandbox with no internet access, and provider credentials are never exposed to them.
Verdicts are signed and independently verifiable, with the evidence behind every finding. Check the signature with the CLI.
Who reviews the pull requests?
A Copilot approval can now count as a required review.
An admin can now let Copilot's approval count toward required approvals.
Most pull requests still get no AI review.
In a 2026 sample of 22,000 developers, three in four pull requests had none.
Source: Faros AI, 2026
Benchmark
Cerebe caught 79% of known issues; the best of five other tools caught 63%.
Serious issues: 82% against 68%. Martian's public Code Review Bench, our own run of the public 50-PR set, September 2026, every tool scored by the same judge. It also comments more; cutting that is current work.
Pricing
Public repositories are reviewed free. Pro starts with a 7-day free trial. Pro and Max bill per developer per month; enterprises can run it in their own cloud, with forward-deployed engineering. The CLI is documented on the install page.
Cerebe Services
Want the product built for you? Our forward-deployed engineers deliver AI-native software on the same platform, at AI speed, with Cerebe's approval on every change.
Install Cerebe
Install the GitHub App on one repository. Your next pull request gets a verdict as soon as it is opened or updated, and so does every one after it.