Headroom cuts AI agent tokens by 60–95%

What happened

Headroom launched a tool that compresses everything an AI agent reads—tool outputs, logs, RAG chunks, files, and conversation history—before sending it to the LLM. It offers a Python/TypeScript library, a proxy, an MCP server, and a wrapper for agents like Claude Code, Cursor, Aider, and Copilot CLI. The system also trims output tokens the model writes back by steering verbosity and dialing down thinking effort on routine steps.

Why it matters

Real workloads show 60–95% token savings (e.g., code search fell from 17,765 to 1,408 tokens, SRE incident debugging from 65,694 to 5,118). Since output tokens on Opus-class models cost 5× input, cutting both dramatically lowers LLM costs. Accuracy is preserved—benchmarks like GSM8K, TruthfulQA, and SQuAD v2 show no degradation or slight improvement (TruthfulQA rose 0.030 points).

What to watch

The tool stores originals locally and is reversible via a retrieval function, so agents can fetch full context on demand. Cross-agent memory deduplicates across Claude, Codex, and Gemini. It runs locally by default (data stays on your machine), and the GitHub Copilot CLI subscription mode routes traffic through the local proxy to intercept and compress OpenAI-compatible requests before they reach GitHub's hosted API.