Your AI coding agent is burning tokens on repetitive garbage—logs, search results, tool spam.

Headroom compresses it all before it hits the model. One command wraps $CLAUDE Code, Codex, Cursor, Copilot, Aider:

headroom wrap claude

Published numbers:
→ Production debugging: 55,957 tokens → 24,340 (57% saved)
→ Codebase exploration: 42% saved
→ GitHub issue triage: 30% saved
→ Accuracy intact on math + tool-use checks

Runs local. Adds ~0.2 ms. Zero data sent out.