Foveance

Foveance — the lossless token codec for LLM context, up to 82% fewer tokens

A lossless token-optimization codec for LLM agents. Cut input tokens up to 82% — nothing dropped.

Your AI agent re-sends its whole history on every turn — the same directory listings, tool outputs and stack traces, over and over. You pay for all of it, every time. Foveance is a real compression codec for that context: like gzip finds repeated bytes, it finds the text that already appeared and replaces it with a short back-reference — removing tokens without removing information. It is exactly reversible (unpack(pack(x)) == x), so nothing is summarised or dropped and every fact survives, and you don't change a line of your app.

On real agent traffic that is 75% fewer input tokens (up to 82% with the template pass), and across five models it answered more accurately than the uncompressed context (0.95 vs 0.90) — stripping the repetition helps the model find the fact.

Get started in 30 seconds

You use a coding agent (Claude Code, Codex, aider, …)

pip install foveance
foveance wrap claude          # or:  foveance wrap -- codex "fix the tests"

It runs your tool exactly as before, just cheaper, and prints how much you saved. Your API key is untouched, nothing is stored.

You write Python

pip install foveance

from foveance import shrink

smaller = shrink(messages, budget=2000)   # your OpenAI-style messages list
# ...send `smaller` to your model instead of `messages`. Same answers, fewer tokens.

Just try it (no API key, no GPU)

pip install foveance
foveance demo

Documentation

  • Usage guide — the proxy, foveance wrap, per-tool recipes, and configuration.
  • Architecture — how the store, predictor, allocator, and controller fit together.
  • Theory — the trajectory rate-distortion framework and the five theorems.
  • Baselines — the policy arms and how they compare.
  • Limitations — the honest failure modes and when the cheap heuristic suffices.
  • Novelty & positioning — what is and isn't claimed as new (prior-art table).

Apache-2.0 licensed.