I kept losing hours in unfamiliar codebases, jumping file to file, guessing what breaks before a refactor. LLMs were great at answering snippet questions but… they guess. They can't see the actual import graph, so they'd confidently draw relations that don't exist. That's the problem RepoGraph AI solves. Open any folder and it indexes itself, no commands, no config. Every edge in the graph is labeled EXTRACTED (found in a real import), INFERRED (0.85) or AMBIGUOUS (0.3), so you always know what to trust. Communities detect themselves (watch backend/frontend split apart at 0:21 in the demo), flow diagrams export as .mmd/.svg/.json, and everything works with zero API keys — the static mode runs fully offline. AI is optional, not the point: bring your own Groq (free tier), Ollama (100% local), Gemini, Claude or OpenAI for summaries and grounded Q&A that cite files and abstain when the graph doesn't support an answer. The piece I'm proudest of: one-click REPOGRAPH-CONTEXT.md export, a token-compact context file that grounds Copilot, Cursor, Claude Code, Codex or OpenCode in your repo's real structure instead of their guesses. I'd love your feedback, especially if you run it on your messiest real-world repo and tell me what the graph gets wrong. That's exactly what the provenance labels are for.

RepoGraph AI

The AI coding-agent graph for any repository just one click

The AI coding-agent graph for any repository just one click