Headroom open-sourced: compresses tool outputs, logs, files, and RAG chunks before reaching the LLM

Headroom offers four integration paths: library, proxy, agent wrapper, and MCP server. It runs locally, works with Python/TypeScript via compress(messages), can proxy any language with zero code changes via headroom proxy --port 8787, and wraps a dozen-plus coding agents (Claude Code, Codex, Cursor, Aider, Cline, etc.) with one command via headroom wrap. Compression uses CacheAligner, ContentRouter, and SmartCrusher/CodeCompressor/Kompress-v2-base for JSON, source code, and text respectively, with original content cached locally for retrieval. In official benchmarks, an SRE incident debugging scenario dropped from 55,957 to 24,340 tokens (57% saved), codebase exploration saved 42%, GitHub issue triage 30%, and code search 21%. Compression takes 0.21ms (10K-token JSON result) to 1.4ms (100K tokens), with no agent latency impact. GSM8K and TruthfulQA accuracy showed no significant difference, and SQuAD v2 and BFCL held at 97% under 19%-32% compression. For developers running local open-source models or using Claude API and OpenAI-compatible endpoints, Headroom can also cut output tokens through the proxy: appending a brief instruction for concise replies and lowering reasoning effort on follow-up turns after tool results, with official reports showing ~31.7% output reduction (95% CI 27.7%-35.7%). GitHub Copilot CLI subscription users can route traffic to the local proxy via headroom copilot-auth login, while VS Code Copilot users keep their existing model selection.