aimaghsoodi/foveance

Foveance is a Python tool that shrinks how much text you send to an AI model without losing any of the information in it. AI coding assistants and chat agents tend to resend their entire conversation history on every single turn, including the same file listings and repeated error messages, and you pay for all of that repeated text each time. Foveance works like a compression tool such as gzip: it finds text that already appeared earlier and replaces the repeat with a short reference back to it, so the model still receives every fact, just without the repetition. According to the README, this approach cuts the number of input tokens by up to 82 percent while staying fully reversible, meaning nothing is summarized or dropped, and the compressed text can be turned back into the exact original. The project's own benchmarks claim it beats other compression methods that are lossy, meaning they can quietly drop facts to save space, and even report that answer accuracy on test questions went up slightly once the repeated text was removed. There are three ways to use it. You can wrap an existing coding agent, such as Claude Code, Codex, Cursor, or aider, with one command, which runs the tool exactly as before but compresses traffic behind the scenes and shows a local dashboard of tokens saved. You can call a single Python function directly in your own code to compress a list of chat messages before sending them to a model. Or you can run a built in demo command to see the tool work on a sample without needing an API key. Installation is a simple pip install, with an optional extra package for machine learning features used in benchmarking. The project is released under the Apache 2.0 license and includes public benchmark scripts and result files so its performance claims can be checked.