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

Описание и применение

Mixdog — открытый агент для программирования, использующий на 63 % меньше токенов, чем Codex CLI, при том же уровне успешности задач.

Задача

Claude and ChatGPT subscriptions hit limits because most tokens were spent on reading files, re‑sending context, and full reasoning for every small step.

Результат

Mixdog achieved the same success rate as Codex CLI while using 63% fewer tokens and 39% lower cost, demonstrating token‑efficient coding workflow.

Рабочий процесс

• 1. Sign in with existing Claude, ChatGPT, Grok or Muse Code subscription or provide an API key. 2. Run Mixdog locally or via npm. 3. Mixdog uses a lightweight local model to judge each message and tool step, selecting minimal reasoning effort. 4. It keeps tool results and history compact and displays a context view. 5. Execute coding tasks; Mixdog logs verdicts, scripts, and token usage for verification.

Доказательства

I built Mixdog, an open‑source coding agent focused on doing the same work with fewer tokens.

Источник

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:

Источник

Total tokens: 156.5M for Mixdog, 421.4M for Codex CLI

Источник

That's 63% fewer tokens at the same success rate, and fewer tokens on 87 of the 89 tasks.

Источник

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.

Источник