headroom

Compress LLM context by applying specialized methods to JSON, code, and long text to reduce token usage while preserving the information needed for accurate downstream reasoning. Content routing automatically detects the input type and selects the appropriate compression strategy.

You can reduce token usage in a proxy workflow by running context compression as a proxy layer. Existing OpenAI-compatible setups benefit from deterministic compression without requiring changes to application logic.

Context compression works with JSON responses, source code, and long tool outputs using specialized methods such as SmartCrusher, CodeCompressor, and LLMLingua-2 to shrink verbose text while retaining reasoning signal.

You need Python 3.10 or later and a headroom virtual environment to set up deterministic LLM context compression. Proxy or API support is also required for integration into conversation workflows.

The best way to shrink long tool outputs is to apply context compression to the latest tool output before sending it back into the LLM context. Structured compression methods target redundant text to reduce tokens while maintaining accuracy.