I built Argus because I noticed a gap: AI agents write code, but then the same AI (or no AI at all) reviews it. That's like grading your own homework.
Argus is a local-first CLI that reviews code with full codebase context — not just the diff. It builds a structural map (tree-sitter + PageRank), runs semantic search over related files, analyzes git history for hotspots, and feeds all of that to the LLM alongside the changes. The result is reviews that actually understand what your code does.
What it does:
- argus review — context-aware AI code review (stdin, local diff, or GitHub PR) - argus map — ranked codebase structure overview - argus search — semantic + keyword hybrid search - argus history — hotspot detection, temporal coupling, bus factor - argus mcp — MCP server for Cursor/Windsurf/Claude Code
Key decisions:
- Multi-provider — OpenAI, Anthropic, or Gemini. Switch in one line. Gemini free tier = zero cost. - One binary — Rust, 10K+ LOC, 264 tests. `cargo install argus-ai` or `npx argus-ai`. - Noise reduction — confidence filtering, dedup, anti-hallucination prompts. We'd rather miss a comment than hallucinate one.
Quick start: `npx argus-ai init && git diff HEAD~1 | npx argus-ai review --repo .` — zero to first review in 60 seconds.
Feedback welcome — especially on review quality vs other tools you've tried.