GitHub added 36 million new developers in 2025. That sounds great until you're the maintainer receiving 200 issues a week, 80% of which say "it doesn't work" with zero context.
In February 2026, an AI agent literally wrote a blog post publicly shaming a Matplotlib maintainer for rejecting its PR. This is where we are now. The inmates are running the asylum.
GitHub calls it "AI slop." The Octoverse report describes it as "a denial of service attack on human attention." Maintainers are burning out faster than ever. Review time has increased faster than the number of people available to do reviews.
I built RepoKeeper to fix this. It's an open source, self-hosted AI agent that handles the boring 80% of repo maintenance automatically.
Here's what it does:
ISSUE TRIAGE: A new issue arrives. RepoKeeper reads it, classifies it (bug, feature, question, docs, invalid), applies labels, and responds contextually. Vague issues get a "needs-more-info" label and specific follow-up questions. Detailed bugs get triaged properly. All in under 3 seconds.
DUPLICATE DETECTION: RepoKeeper uses AI semantic comparison to detect when someone opens "app won't start" three days after someone else opened "nothing works after install." It flags duplicates and links them. No more answering the same question five times.
PR SUMMARISATION: Large PRs are intimidating. RepoKeeper generates plain-English summaries with per-file descriptions, size labels, and breaking change flags. Your reviewers see context before they see code.
AI CODE REVIEW: Line-by-line review comments that actually make sense. Test gap detection. Configurable focus areas. It remembers what patterns you've already accepted so it doesn't repeat itself.
RELEASE NOTES: When a PR merges, RepoKeeper auto-generates changelog entries. No more "what shipped this week?" meetings.
The best part? It runs on YOUR machine. No SaaS. No vendor lock-in. Your code never leaves your infrastructure. Drop it on a $5 VPS and forget about it.
One instance can manage multiple repos. Define per-repo overrides for triage, code review focus areas, even different AI models per repo. Docker Compose with Nginx and automatic HTTPS included. Or use the systemd service file. Production-ready from day one.
GitHub launched their own "Agentic Workflows" in February. It's locked to GitHub Actions, costs $0.002/minute in compute, requires YAML configuration, and is still in technical preview. RepoKeeper works today, costs nothing, and runs on GitHub, GitLab, Gitea, or anywhere else.
The tech stack is boring on purpose: TypeScript, Express, Octokit. Bring your own AI (Claude, GPT, Ollama). 94 tests passing. CI green. Setup takes 60 seconds with the interactive wizard.
Here's what a real triage looks like. Someone opens an issue titled "broken" with body "doesnt work help." Within 2 seconds, RepoKeeper:
- Detects it's a vague report (no error messages, no steps to reproduce, no OS/version info)
- Applies the "needs-more-info" label
- Posts a contextual comment asking for specific details
- Checks existing issues for semantic duplicates
No human touched it. The maintainer wakes up to a clean, labelled, responded-to issue queue.
The numbers that matter for maintainers: GitHub saw 1 billion commits in 2025 (up 25% year over year). 43.2 million pull requests merged monthly. The ratio of contributors to maintainers is getting worse, not better. Tools like RepoKeeper aren't a nice-to-have anymore. They're survival infrastructure.
I built this in a week using Claude as my coding partner. The irony isn't lost on me: using AI to build a tool that protects maintainers from bad AI contributions. But that's exactly the point. AI should serve the maintainer, not the contributor.
RepoKeeper is live, public, and free: github.com/GodsBoy/repokeeper
Star it if you're tired of "it doesn't work" issues. Fork it if you want to make it better. The code is MIT licensed because the best tools should be free.
Top comments (0)