Draft pull request reviews for you to edit and send
Fifteen PRs waiting for my review had become a normal morning. Task: Fifteen PRs waiting for my review had become a normal morning. Delegating to Claude Code helped, but I still had to start each review manually. Having agents post directly to GitHub created another problem: comments I hadn't checked became someone else's work to deal with. So I built Cerber. It watches GitHub for PRs requesting your review, runs Claude Code over them, and prepares drafts in a local web UI. You get a summary, a walkthrough of the changes, and proposed inline comments. You can discuss a finding with the agent, rewrite it or delete it, then send the review under your own account. A few implementation details: - It runs through your existing Claude Code CLI login. - Drafts and review state stay on your machine; code is sent to Anthropic. - Reviews are read-only by default. Command execution requires explicitly trusting the author, their GitHub team or organization. - Sending is manual by default. There's an opt-in auto-send mode for approvals meeting a confidence threshold. Cerber is MIT-licensed and requires an existing paid Claude Code subscription. GitHub: https://github.com/fullstackhouse/cerber Full story: https://fullstack.house/blog/cerber How are you handling your review queue? Did anybody of you build such a tool? I wish a feature like that be built into GitHub. Copilot does it a bit... but the general GH's UX doesn't feel "right" yet, too noisy for me.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include a paid Claude Code subscription and GitHub access. The tool is MIT-licensed. Code is sent to Anthropic, while drafts remain local. Command execution is restricted by default, requiring explicit trust for authors or organizations. Auto-send is opt-in and relies on a confidence threshold. Source reliability is low as the original post is inaccessible (HTTP 403), and details are derived from secondary summaries..
Tools
Published on Agent Work Atlas:
Overview
Fifteen PRs waiting for my review had become a normal morning.
Workflow
Fifteen PRs waiting for my review had become a normal morning. Delegating to Claude Code helped, but I still had to start each review manually. Having agents post directly to GitHub created another problem: comments I hadn't checked became someone else's work to deal with. So I built Cerber. It watches GitHub for PRs requesting your review, runs Claude Code over them, and prepares drafts in a local web UI. You get a summary, a walkthrough of the changes, and proposed inline comments. You can discuss a finding with the agent, rewrite it or delete it, then send the review under your own account. A few implementation details: - It runs through your existing Claude Code CLI login. - Drafts and review state stay on your machine; code is sent to Anthropic. - Reviews are read-only by default. Command execution requires explicitly trusting the author, their GitHub team or organization. - Sending is manual by default. There's an opt-in auto-send mode for approvals meeting a confidence threshold. Cerber is MIT-licensed and requires an existing paid Claude Code subscription. GitHub: https://github.com/fullstackhouse/cerber Full story: https://fullstack.house/blog/cerber How are you handling your review queue? Did anybody of you build such a tool? I wish a feature like that be built into GitHub. Copilot does it a bit... but the general GH's UX doesn't feel "right" yet, too noisy for me.
Reported result
No measured outcomes or success metrics are reported in the source. The author describes the workflow as a solution to manual review initiation and unverified agent comments, but provides no data on time saved or review quality improvements. A measurable acceptance check would be tracking the reduction in time from PR notification to review submission compared to the manual baseline.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include a paid Claude Code subscription and GitHub access. The tool is MIT-licensed. Code is sent to Anthropic, while drafts remain local. Command execution is restricted by default, requiring explicit trust for authors or organizations. Auto-send is opt-in and relies on a confidence threshold. Source reliability is low as the original post is inaccessible (HTTP 403), and details are derived from secondary summaries.
Try it
Act as a developer building a PR review assistant. I need you to create a local web UI tool called 'Cerber' that integrates with GitHub and Claude Code. Inputs: [GitHub Repo URL], [Claude Code API Key], [Review Confidence Threshold]. Steps: 1. Set up a watcher for PRs requesting my review in [GitHub Repo URL]. 2. For each new PR, trigger Claude Code to analyze the diff and generate: a summary, a walkthrough of changes, and proposed inline comments. 3. Display these drafts in a local web UI. Do not post to GitHub automatically. 4. Allow me to edit, delete, or discuss each comment with the agent. 5. Provide a 'Send Review' button that posts the final approved comments under my account. 6. Implement a safety check: only enable auto-send for approvals if the confidence score exceeds [Review Confidence Threshold], otherwise require manual confirmation. Deliverables: A working prototype script and UI mockup. Quality Check: Run a dry test on one sample PR. Show me the generated draft and the UI state before attempting any GitHub API calls. Confirm with me before proceeding to full implementation.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.