Put a Claude-Powered Code Reviewer in Your Pull Requests
You know that feeling when you push a PR, and you're pretty sure the linter is going to complain about something you missed? Or when you want a second pair of eyes on a diff, but your team is busy and the CI queue is backed up? OpenReview is a self-hosted bot that slides into that gap—you mention it in a PR comment, and it spins up a sandbox, runs your tooling, and posts suggestions directly on the lines that need fixing.
It's an open-source project from Vercel Labs, and it's built to be deployed on Vercel's infrastructure. The bot uses Claude Sonnet 4.6 through the AI SDK, runs everything in isolated Vercel Sandboxes, and orchestrates the whole flow with Vercel Workflow. Yes, it's deeply tied to Vercel's platform—but that's also what makes it surprisingly easy to get running.
What It Does
OpenReview is a GitHub App that listens for webhook events. When you comment @openreview on a pull request (optionally with specific instructions like "check for memory leaks" or "run the formatter"), the bot kicks off a workflow.
Here's the flow: the webhook handler starts a Vercel Workflow, which checks if the bot has push access to the repo, then creates a sandboxed environment. Inside that sandbox, it clones your repo on the PR branch, installs dependencies, and configures git. Then a Claude-powered agent takes over—it reads the diff, explores the codebase, and runs linters, formatters, and tests as needed.
The agent posts its findings as PR comments with inline suggestions. These aren't just text comments—they're GitHub suggestion blocks, so you can apply the fix with a single click. If the agent makes changes itself (say, fixing a formatting issue), it commits and pushes those changes directly to your PR branch. Finally, the sandbox is cleaned up.
You can also react to suggestions with 👍 or ❤️ to approve them, or 👎 or 😕 to skip them. That reaction triggers another workflow run, which presumably tells the agent to apply or discard the suggestion.
Why It's Cool
The "mention to trigger" model is the killer feature here. You don't need a separate dashboard, a Slack integration, or a scheduled job. You just tag the bot in a comment, and it runs. That's a low-friction interaction that fits how developers already talk about PRs.
The sandboxed execution is another strong point. OpenReview doesn't just read your code statically—it actually runs your project's tooling. That means it can catch issues that a pure LLM code review would miss, like build failures or test failures. And because it has push access, it can fix simple bugs and formatting issues without you having to copy-paste anything.
The reaction-based approval system is a clever touch. Instead of asking the bot to make changes and then manually verifying them, you can approve or reject each suggestion with a single emoji reaction. It's a lightweight feedback loop that keeps you in the GitHub UI.
It's also extensible. The bot ships with built-in review "skills," but you can add custom ones by dropping files into .agents/skills/ in your repo. That means you can teach it your team's specific conventions or run custom checks beyond what's pre-configured.
One honest caveat: the README explicitly says this is in beta and was built as an internal project for Vercel's team. Expect rough edges and breaking changes. This isn't a polished commercial product—it's a well-architected internal tool that's been open-sourced.
How to Try It
The setup involves three steps, and the README makes it fairly direct.
First, deploy the project to Vercel. There's a one-click deploy button in the README that handles the environment variables for you, or you can clone the repo and deploy it manually.
Second, create a GitHub App. You'll need to set up a new GitHub App and configure the webhook to point at your deployed instance.
Third, connect the pieces. The deploy process will ask for GITHUB_APP_ID, GITHUB_APP_INSTALLATION_ID, GITHUB_APP_PRIVATE_KEY, and GITHUB_APP_WEBHOOK_SECRET. It also integrates with Upstash KV for storage, which gets provisioned as part of the Vercel deployment.
Once that's set up, you just install the GitHub App on your repos and comment @openreview on a PR to trigger a review. The full setup details are in the repository at github.com/vercel-labs/openreview.
Final Thoughts
OpenReview is best suited for developers who are already comfortable with Vercel's ecosystem and want a self-hosted alternative to paid code review tools. It's not a drop-in replacement for a human reviewer—it's more like an automated assistant that handles the mechanical parts of code review: running linters, catching formatting issues, and fixing simple bugs.
The architecture is clean—separating the webhook handler, the durable workflow, and the sandboxed execution is a solid design that makes the system resumable and reliable. If you're curious about how Vercel's own tools fit together, this project is also a great reference implementation.
It's in beta, so you might hit rough edges. But for a free, self-hosted AI reviewer that actually runs your code, it's worth a weekend experiment.
Follow @githubprojects for more developer tools and open source projects.