I built an open-source coding agent that uses 63% fewer tokens than Codex CLI

TRIBGAMES DEV — Sat, 24 Oct 2026 14:30:53 +0000

My Claude and ChatGPT subscriptions kept hitting their limits, and most of the tokens weren't going into the actual fix. They went into reading files, re-sending context, and running every small step at full reasoning effort. So I built Mixdog, an open-source coding agent focused on doing the same work with fewer tokens.

The numbers

I ran all 89 Terminal-Bench 2.1 tasks, 5 trials each, with the same model (GPT-5.6 Sol, xhigh) on both sides, scored by the official Harbor verifier:

	Mixdog	Codex CLI
Solved	385 / 445 (86.5%)	383 / 445 (86.1%)
Total tokens	156.5M	421.4M
Cost	39% lower

That's 63% fewer tokens at the same success rate, and fewer tokens on 87 of the 89 tasks. With Claude Opus 5, Mixdog solved 2 more tasks than Claude Code at 19% lower cost. Every verdict, log and script is published in the repo, so you can check or rerun it.

Where the savings come from

A lead and a team. A strong model plans; cheaper models search, edit and review in parallel. You can scale from a single model to a full swarm and switch mid-conversation.
Auto reasoning effort. A 150 MB model running locally on the CPU judges each message and tool step in about 0.1 s and picks the effort it needs, so simple steps don't burn xhigh reasoning.
Lean context. Tool results and history are kept compact, and a context view shows exactly where every token goes.

Use what you already pay for

Sign in with the Claude, ChatGPT, Grok or Muse Code subscription you already have, paste an API key, or run local models on your own GPU. A usage panel shows quota left and reset times for each plan.

Beyond coding

Mixdog is a full desktop workspace: editor, terminals, Git and GitHub, code graph and language servers. It can also control the browser and desktop, work with Office and PDF files, generate images and video, keep working on goals, schedules and webhooks, and continue from your phone with end-to-end encryption.

It runs on Windows, macOS and Linux with no config files; a five-step tutorial gets you started. It's free and Apache-2.0.

Try it

Download and source
Benchmark details
Or from the terminal: npm install -g mixdog

I'd love feedback, especially on where it still wastes tokens in your own workflow.