I open-sourced Tokenade, my token optimizer for coding agents. Here is why the paid model did not hold

Tokenade is free and open source as of version 2.0, under MIT OR Apache-2.0. The code is on GitHub: https://github.com/pi-infected/tokenade

I tried to sell it first. Here is why that did not hold, what Tokenade does, and how to check it on your own sessions.

Why the paid model did not hold

Most people run Claude Code, Cursor or Codex on a subscription. They never see a token bill, so a tool that shrinks the bill solves a problem they do not feel. The people who do pay per token call the API directly, and for them the useful version is a proxy in front of the provider. I do not want to run a proxy business.

So the honest move was to give it away and let the code speak.

What it does

A coding agent pays for what it reads, and it reads the same things again with each turn: build logs, test output, whole files it opened once and kept in the history. Tokenade works on that layer, on your machine.

• Command output. When the agent runs a noisy command, Tokenade compacts what comes back before the model reads it. tokenade raw <cmd> bypasses it when you want the full output.
• Finding code without reading it. tokenade map, query <symbol>, skeleton <file> and semantic "<question>" give the agent the structure it needs instead of whole files.
• Documents and pages. tokenade read turns a PDF, a spreadsheet or a long file into the text that matters. tokenade web and tokenade search do the same for web pages and search results.
• API users. tokenade llm-proxy runs locally between the agent and the provider. It folds tool results already in the history and sends an identical tool output only once. Run it, or fork it.

tokenade install detects the agents already on the machine and wires itself in. There is nothing to configure.

The number

On THOL, a public benchmark of token-harness optimizers that I maintain, Tokenade reduces token costs by 39% on average on long sessions (over 200k tokens) with Claude Code, measured on 2026-09-14. That is an average, not a ceiling, and the raw runs are public: https://pi-infected.github.io/token-harness-optimizer-leaderboard/

Try it

npm install -g @tokenade/cli
tokenade install

After a long session, tokenade dashboard shows what it saved, graphed by token category and by model.

If it saves you nothing, I want to know why: open an issue with the command or the setup. If it helps, a star on the repo is the quickest way to tell me.