_{macOS button is for Apple Silicon (M1 and later). Intel Mac? Download for Intel Mac. Linux button is the x64 AppImage; .deb and ARM64 builds are listed below.}

Website · Benchmarks · Docs

Highlights

63% fewer tokens than Codex CLI on the same model, at the same success rate.
Auto reasoning effort — each message and tool step gets the effort it needs, judged in about 0.1 s by a 150 MB model that runs locally on the CPU. On by default.
Use the Claude, ChatGPT, Grok, or Muse Code subscription you already have, API keys, or local models on your own GPU.
Mix models by role and scale from Solo to Swarm — a strong model leads while cheaper ones search, edit, and review in parallel. Switch the workflow or mode mid-conversation, and set fallback models for usage limits.
Bots and rooms — your own bots with their own models, tools, and permissions, group rooms where they talk together, and a representative bot that sits above the composer on your desktop.
A full workspace: editor, terminals, Git and GitHub, code graph, and language servers, with themes and photo skins.
Keeps working on its own with Goals, schedules, and webhooks — and continues from your phone, end-to-end encrypted.
Goes beyond code with browser and desktop control, Office and PDF documents, and image and video generation.
Runs on Windows, macOS, and Linux. No config files.

Installation

Download the desktop app:

Platform	Download
Windows (x64)	`mixdog-desktop-win-x64.exe`
macOS (Apple Silicon)	`mixdog-desktop-mac-arm64.dmg`
macOS (Intel)	`mixdog-desktop-mac-x64.dmg`
Linux (x64, AppImage)	`mixdog-desktop-linux-x86_64.AppImage`
Linux (x64, .deb)	`mixdog-desktop-linux-amd64.deb`
Linux (ARM64, AppImage)	`mixdog-desktop-linux-arm64.AppImage`
Linux (ARM64, .deb)	`mixdog-desktop-linux-arm64.deb`

Then sign in with a ChatGPT, Claude, Grok, or Muse Code account, or paste an API key. A five-step tutorial walks you through the rest.

[!NOTE] The Windows installer and app are code-signed, and the Mac app is signed and notarized. If SmartScreen still appears, choose More info → Run anyway.

Or run it in the terminal (Node.js 22.19+ or 24+):

npm install -g mixdog
mixdog

Supported providers

Anthropic API keys and Claude account OAuth
OpenAI API keys and ChatGPT/Codex account OAuth
Google Gemini API keys
xAI API keys and Grok account OAuth
Muse Code (Meta) account OAuth
OpenRouter, DeepSeek, and OpenCode Go
Built-in Local Provider — download and run models in the app (Windows x64 with an NVIDIA GPU)

Cursor, Antigravity (Gemini), and MiMo OAuth are off by default under Settings → Developer; using them through OAuth risks account restrictions.

Headless mixdog exec

mixdog exec runs one non-interactive, single-model session without personal memory, prior sessions, skills, MCP servers, or plugins:

mixdog exec --provider openai-oauth --model gpt-5.6-sol --effort xhigh "fix the failing test"

Web search is off by default (--web-search enables it). This does not block shell networking — headless exec is not an offline sandbox. Run mixdog --help for the full reference.

Benchmarks

Bar chart: Mixdog vs Codex CLI on Terminal-Bench 2.1

All 89 Terminal-Bench 2.1 tasks × 5 trials on GPT-5.6 Sol xhigh, scored by the official Harbor verifier.

Mixdog used fewer tokens on 87 of 89 tasks. With Claude Opus 5, it solved 2 more tasks than Claude Code at 19% lower cost. Every verdict, log, and script is published in benchmarks/terminal-bench-2.1/, and Context efficiency explains where the savings come from.

Benchmark history

Newest first. Each run links to its details, records, and caveats.

Run	Model · trials	vs	Score	Cost	Final context	Time / trial
2026-10-15	GPT-6.1 Sol xhigh · `k=1`	Codex CLI	77 vs 78 /89	−55%	−68% (10.8K vs 33.6K)	344s vs 380s
2026-08-25	GPT-5.6 Sol xhigh · `k=5`	Codex CLI	385 vs 383 /445	−39%	−46% (18.5K vs 34.3K)	415s vs 437s
2026-08-25	Claude Opus 5 · `k=1`	Claude Code	79 vs 77 /89	−19%	−28% (27.6K vs 38.2K)	610s vs 708s

The 2026-10-15 run is the first full run with the effort judge active on OpenAI routes, with web search on for both sides. It is a single pass, so the k=5 chart above stays the reference.