Headroom

A lot of what your assistant reads is filler. Records of what programs printed, long stretches of data, repeated boilerplate. Every word of it uses up tokens, the units AI reads and bills by, and almost none of it makes the answer better. Headroom compresses that material before it reaches the model. The project measures around 20 percent fewer tokens on coding work and between 60 and 95 percent on data files, with the same answers coming back.

Work that pushes a lot of material through the assistant: reading logs, handling large data files, or long runs where the assistant's working memory fills up and earlier detail starts falling out.

1. Run the command below. It needs Python 3.10 or newer; typing python3 --version in the Terminal tells you whether you have it. If it is missing, Python is a free install from python.org, and pip, Python's own installer, arrives with it.
2. Follow its guide, on the project page linked in the sidebar, to add it to your assistant's connected tools. That is the simplest of its ways to run, and the guide covers the others.
3. Carry on as normal. It compresses in the middle and the answers come back unchanged.