DEV Community

Kim
Kim

Posted on

My AI Agent Kept Recommending Abandoned Repos. So I Built RepoCrunch.

#github

I've been building AI agents that evaluate GitHub repos — picking dependencies, reviewing codebases, comparing tools. The problem? They hallucinate.

Ask an LLM "is this repo well-maintained?" and you'll get a confident answer based on vibes. Not data. My agent kept recommending packages that were archived, had 2 commits in the last year, or had open CVEs.

So I built RepoCrunch — a CLI tool and MCP server that gives AI agents ground-truth GitHub intelligence.

What it does

Point it at any public GitHub repo and get deterministic, structured JSON:

  • Tech stack detection — languages, frameworks, build tools
  • Dependency analysis — what it uses, version info
  • Health metrics — stars, issues, contributors, commit frequency
  • Security indicators — CI/CD setup, known vulnerability patterns
  • Architecture signals — project structure, patterns used

The key word is deterministic. Same repo, same output. No interpretation, no hallucination.

Three ways to use it

1. CLI 

pip install repocrunch
repocrunch analyze https://github.com/fastapi/fastapi
Enter fullscreen mode Exit fullscreen mode

2. MCP Server (for AI agents) 

repocrunch mcp
Enter fullscreen mode Exit fullscreen mode

Drop this into your Claude Desktop, Cursor, or any MCP-compatible client. Your agent gets real repo data instead of guessing.

3. REST API 

repocrunch serve
Enter fullscreen mode Exit fullscreen mode

Why MCP matters here

The Model Context Protocol lets AI agents call external tools. When your agent needs to evaluate a dependency, it can call RepoCrunch and get facts — not generate them from training data.

This is the difference between:

  • ❌ "FastAPI appears to be well-maintained based on my knowledge"
  • ✅ "FastAPI has 847 contributors, 12.3k commits, last push 2 hours ago, MIT license, CI passing"

The real use case

I built this for my own AI agent workflow. When my agent evaluates whether to use a library, it runs RepoCrunch first. The structured output feeds directly into the decision — no guessing.

But it works for anyone who needs repo intelligence:

  • Dependency auditing — are your deps actually maintained?
  • Due diligence — evaluating repos before adoption
  • Portfolio monitoring — tracking health of repos you depend on

Try it 

pip install repocrunch
repocrunch analyze https://github.com/your-favorite-repo
Enter fullscreen mode Exit fullscreen mode

GitHub: github.com/kimwwk/repocrunch
PyPI: pypi.org/project/repocrunch

Open source, MIT licensed. Feedback welcome — especially from folks building MCP toolchains.

Built this as a solo dev scratching my own itch. If you're doing anything with AI agents + code evaluation, I'd love to hear how you're solving the hallucination problem.

Top comments (0)