Back home

Mixdog

x+web+github 0.66
Visit source

Intro

Mixdog is a free, open-source (Apache-2.0) desktop coding agent for Windows, macOS and Linux, built by the Korean studio tribgames (트리비게임즈). Its tagline: "The open-source coding agent that uses 63% fewer tokens."

The core idea is a lead-and-team workflow: a strong model plans, while cheaper models search, edit and review in parallel. You bring the subscription you already have — Claude, ChatGPT/Codex, Grok or Muse Code accounts — or API keys (OpenAI, Anthropic, Gemini, xAI, OpenRouter, DeepSeek) or local models.

Repo: github.com/tribgames/mixdog · Product page: tribgames.com/products/mixdog · Launch write-up: dev.to.

Explain

Key points

  1. Token efficiency is the headline: on all 89 Terminal-Bench 2.1 tasks × 5 trials (445 per agent), same model (GPT-5.6 Sol xhigh) on both sides and scored by the official Harbor verifier, Mixdog used 156.5M tokens vs 421.4M for Codex CLI (−63%) with 86.5% vs 86.1% success, and fewer tokens on 87 of 89 tasks. With Claude Opus 5 it solved 2 more tasks than Claude Code at 19% lower cost. Logs, verdicts and scripts are published in the repo (vendor-run benchmark).
  2. Where savings come from: a lighter first request (≈4.7k vs 14.4k tokens of instructions/tool specs), fewer round trips (~10 vs 16 requests per trial), filtered tool output (≈95% of terminal output dropped before reaching the model) and shorter replies.
  3. Auto reasoning effort: a ~150 MB local model on the CPU judges every message and tool step in ~0.1 s and picks the reasoning effort, so trivial steps don't burn xhigh reasoning.
  4. Solo to Swarm: mix models by role, switch workflow mid-conversation, set fallback models for usage limits.
  5. Full workspace: editor, terminals, Git/GitHub, code graph, language servers; bots and group rooms; goals, schedules and webhooks; end-to-end-encrypted continuation from your phone; browser/desktop control, Office/PDF documents, image and video generation.

Why this radar term

The repo was created in June 2026, but v1.4.0 and v1.4.1 shipped on 2026-10-27 and Mixdog surfaced in Chinese-language X open-source roundups in this window (alongside the Oct 24 dev.to launch post). It is a distinctive, newly visible product name in the crowded coding-agent space.

Usage

Get started

  1. Desktop app: download the installer for your platform from the GitHub releases (Windows x64, macOS Apple Silicon/Intel, Linux x64/ARM64 AppImage or .deb).
  2. CLI (Node.js 22.19+ or 24+):
npm install -g mixdog
mixdog
  1. Sign in with a ChatGPT, Claude, Grok or Muse Code account, or paste an API key; a five-step tutorial covers the rest.
  2. Pick a lead model for planning and cheaper models for search/edit/review, and keep auto reasoning effort on (default).

Tips

  • You pay providers directly; the app itself is free.
  • On Intel Macs, auto reasoning effort and semantic memory search are not available.
  • Treat the 63% figure as a vendor benchmark — rerun the published scripts on your own tasks before switching.