A library, proxy, and MCP server for local, reversible compression of agent context
A library, local proxy, agent wrapper, and MCP server that compresses messages, tool outputs, JSON, code, and RAG chunks. Its distinguishing feature is reversible compression (CCR): originals are stored locally and can be retrieved when the model needs them. The benchmark page honestly distinguishes substantial gains on repetitive JSON/log payloads from more modest gains on plain text [8]. It can be integrated with your own Python/TypeScript agent or an existing Claude/Codex/Cursor session.
Technical teams that want to integrate compression into their own agent or provider proxy while keeping data local.
Requires Python 3.13 and model downloads, so it isn't lightweight. Since the proxy can affect all provider traffic, it should also be reviewed separately in enterprise networks. Savings vary with your traffic; run paired measurements.
uv tool install --python 3.13 "headroom-ai[all]" headroom wrap claude