Headroom: How Netflix Cut AI Agent Token Costs by 92%
Tejas Chopra (Netflix) was burning $200/day on tokens with his AI agents. Not a model problem — pure context volume. Every tool call, log, and search result accumulated until the window became unmanageable. His solution: Headroom, an open source compression layer between your agent and the LLM.
Headroom compresses context using six methods by type: AST for code (tree-sitter), SmartCrusher for JSON, and a HuggingFace model for prose. Real results: 17,765 → 1,408 tokens in code search (92%). Cost: from $200 to $30/day. No degradation on accuracy benchmarks.
Install and activate with: npm install -g headroom-ai && headroom wrap claude