Player not loading? Watch on YouTube
Headroom is an open source tool that compresses tool call output before it reaches an LLM. The tutorial explains its Python proxy and demonstrates a Claude Code setup. Compression happens locally, but the example still sends requests to Anthropic's servers.
The speaker describes different compressors for JSON arrays, syntax trees and build logs. A local model handles plain text. Compressed output includes a hash that lets the model request the original data. For installation, the presenter uses uv with Python 3.12 and says newer Python versions do not work with the demonstrated setup. The walkthrough also covers the Python SDK, then enables the ML and code-aware options and points Claude Code's base URL at the proxy.
A synthetic log example reports 98% fewer tokens, saving over 17,000 tokens after summarizing 419 similar informational logs. Claude initially says it lacks enough information; another run gives a fuller answer. These are demonstration results, not a guarantee of equivalent answers or savings on every task.
A second example asks the coding assistant to explain TypeScript files across five packages. The presenter reports no savings at low Opus effort, with savings appearing at medium effort. Requests for original data add a round trip and can increase token use. The closing comparison describes Caveman as shortening model responses, while Headroom compresses inputs; the speaker suggests using both together.
