chopratejas/headroom

Automatically compresses boilerplate from tool outputs, database queries, RAG retrievals, and file reads (typically 70-95% of context) before sending to the LLM, reducing token usage while preserving accuracy. Available as a Python/TypeScript SDK function, transparent HTTP proxy, or framework integrations for LangChain, LiteLLM, Agno, and coding agents (Claude Code, Cursor, Aider). Uses statistical anomaly detection and adaptive sampling to preserve critical information—evidenced by 87-92% token savings on production workloads with no accuracy loss.

724 stars. Actively maintained with 160 commits in the last 30 days.