Open-source library, proxy and MCP server that compresses tool outputs, logs, RAG chunks and history before they reach an LLM.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
Headroom (headroom-ai on PyPI and npm) is an Apache-2.0 context-compression layer for LLM applications and coding agents. It compresses tool outputs, logs, files, RAG chunks and conversation history before they are sent to the model, with the aim of returning the same answers using fewer input tokens. Compression runs locally, and originals are cached so the model can retrieve them on demand (CCR).
A content router selects a compressor by content type: SmartCrusher for JSON, CodeCompressor (AST-based) for source code, and Kompress-v2-base for text. The README reports 21–57% fewer input tokens across four agent scenarios, 90%+ on repeated JSON arrays and log lines, and little on prose. It also includes cross-agent memory, output-token reduction, and headroom learn, which mines failed sessions and writes corrections to files such as CLAUDE.local.md, CLAUDE.md or AGENTS.md.
It can be used inline (compress(messages, model=...) in Python or TypeScript), as a proxy (headroom proxy --port 8787), as an MCP server, or by wrapping coding agents such as Claude Code, Codex, Cursor, Copilot, Aider and OpenCode with headroom wrap. A YouTube reviewer noted that it installs Serena by default and that telemetry is on by default. Savings depend heavily on content type.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
An open-source tool that compresses tool outputs, logs, RAG chunks and history before they reach the LLM, claiming 47-92% token savings with preserved accuracy. Works with Claude Code, Cursor and Codex; includes a perf command and 'headroom learn' to write corrections to Claude.md/agents.md. Caveats: installs Serena by default and telemetry is on by default.
Closest neighbours by meaning, not just by category.
Model · Meta
Closed-source Meta model for coding and agentic workflows with 1M-token context and multimodal input, offered via API.
Library · DeusData
Local MCP server that indexes code, docs and infrastructure into a knowledge graph for AI coding agents.
Model · Alibaba
Alibaba's upgraded Qwen3.8-Max snapshot with 1M context, thinking mode and stronger coding and agent performance, via API.
Library
Open Markdown format for giving coding agents project instructions, like a README for agents.
Model
35B Mixture-of-Experts open-weights model from InternScience for long-horizon agentic search, engineering and tool use.
Model · Recursive Intelligence
Dense 65B open-weight language model from Ricursive Intelligence, trained for agentic coding with up to 1M context.
