Headroom

Context optimization layer for LLM apps: compresses what agents read (tool outputs, logs, RAG, file reads) while preserving accuracy.

Headroom docs

Overview

Headroom strips boilerplate and keeps signal — losslessly and locally — before context reaches the model. It targets high-volume agent inputs like JSON tool outputs, build/test logs, search results, diffs, and source code.

How it fits (one line)

Agent/App → (tool outputs, logs, DB reads, RAG chunks, files) → Headroom → LLM provider.

Notable features

  • Wrap existing coding agents: headroom wrap claude|codex|cursor|aider|copilot
  • Proxy mode: run as a local proxy for zero-code integration
  • Content-aware compressors: JSON, code (AST-aware), logs, text, diffs, search results, images
  • CCR (reversible / lossless compression): compress aggressively but allow retrieval of originals
  • Local-first: no data egress; millisecond compression latency
  • Cross-agent memory + learning: mines failed sessions and can write corrections to CLAUDE.md / AGENTS.md / GEMINI.md

Numbers (as marketed)

  • ~87% token reduction on example log debugging (10,144 → 1,260) with identical answer.