headroom: Compress Everything Before It Reaches the LLM, 60-95% Smaller

If you have ever watched an AI agent blow through its context window by dumping raw API responses, file contents, and logs all at once, headroom is the fix. It is an open-source context compression system that processes everything an LLM is about to read—tool outputs, log files, RAG chunks, conversation history—before it actually reaches the model. Trending on GitHub daily with 1,266 new stars today, total 6,148.

What is technically interesting is the content-aware approach. It does not apply one compression algorithm to everything—it runs each content type through a specialized pipeline. SmartCrusher handles JSON structures. CodeCompressor uses AST-aware techniques for code. Kompress-base is a custom model trained specifically on agent traces. CacheAligner optimizes prefix stability for provider KV caches, which matters if you are trying to hit cache on long system prompts.

Three deployment options: library for inline use, proxy server requiring zero code changes, or MCP server for compatible clients. Works with Claude, Cursor, Codex, Copilot, and any OpenAI-compatible stack. There is also a headroom learn mode that mines failed agent sessions and writes corrections to prevent repeating the same mistakes. https://github.com/chopratejas/headroom