I built SigMap in April. By May 5 it had 93 stars — growing slowly through organic
discovery. Then last week happened.
61K Reddit views. 90.5% upvote ratio. 307 new stars in 8 days.
SigMap is now in the top 1.5% of all coding AI repositories on GitHub.
Here's what it does and why the numbers moved.
The problem it solves
When you use Claude Code, Cursor, or Copilot on a large codebase, the tool
reads your open files or sends the whole project. On real repos, that's
80,000+ tokens per session. 13 of 18 repos I benchmarked overflow GPT-4o's
context window entirely.
The AI is working with an incomplete, random picture of your codebase.
How SigMap fixes it
Instead of sending full source files, SigMap:
- Extracts function signatures + import graph from your entire codebase
- Ranks them by TF-IDF against your specific query
- Sends 200–4,000 query-ranked tokens instead of 80,000 random ones
Different question = different context. The AI sees what's relevant, not what's open.
The benchmark numbers
Tested across 18 repos, 90 tasks, no LLM API calls:
- 80.0% hit@5 vs 13.6% random baseline — 5.9× lift
- 96.8% average token reduction
- 52.2% task success vs 10% without SigMap
- 41.0% fewer prompts per task
The 405-repo large-scale token benchmark is on Zenodo (DOI: 10.5281/zenodo.19898842).
Peer-reviewed. Reproducible.
Why zero dependencies matters
Most context tools require a vector database, an embedding model, or cloud
infrastructure. SigMap uses Node.js stdlib only — crypto, fs, path, child_process.
Zero npm dependencies. You can audit every line.
This is a deliberate constraint. A tool that reduces your AI's dependencies
shouldn't add its own.
How to try it
npx sigmap
sigmap ask "how does the auth flow work"
That's it. Works offline. Works with any model.
MCP integration for Claude Code, Cursor, Copilot, Windsurf, Codex, Gemini CLI.
→ GitHub: https://github.com/manojmallick/sigmap
→ Benchmark: https://github.com/manojmallick/sigmap-benchmark-suite
→ Docs: https://manojmallick.github.io/sigmap
Built in Amsterdam. Open source. MIT licensed.
Top comments (0)