Same results.
63% fewer tokens.
Mixdog is a free, open-source coding agent. Connect your existing Claude, ChatGPT, Grok, or Muse Code subscription with no extra fees.
Same model, same score.
Less than half the tokens.
Tested across all 89 Terminal-Bench 2.1 tasks with 445 trials per agent, both running GPT-5.6 Sol xhigh.
View method and raw data →Total tokens (including cached input) · Terminal-Bench 2.1
- 86.5%Success rateCodex CLI 86.1%
- $0.476Cost per trialCodex CLI $0.782
- 87 of 89Tasks with fewer tokens
Scored with the official Harbor verifier using unmodified timeouts and zero retries. Compared to Claude Code on Claude Opus 5, Mixdog cost 19% less and solved 2 more tasks.
How we stop token leaks
- 67%
Lighter first request
System instructions and tool descriptions include only what the model needs: 4.7k tokens vs. 14.4k for Codex CLI.
- 33%
Fewer round trips
Selects the right tools upfront and batches independent steps: only 10 requests per trial vs. 16 for Codex CLI.
- 46%
Filtered output
Tools return only what was requested, and 95% of terminal output is filtered out before reaching the model.
- 8%
Shorter responses
Reduces filler and repetition in model replies.
Everything in one window
Ask once, get it fixed.
Finds failing tests, fixes the bugs, and reruns the suite. Three bugs fixed in 12 seconds.
From solo to swarm.
A strong model plans the work, while cheaper models search, edit, and review in parallel.
Run tasks side by side.
Split the window and assign different tasks to completely independent sessions.
Editor, terminal, Git, and GitHub built in.
Edit code with language server support, preview and edit Markdown, CSV, PDF, and Office files, run tests, commit, and open PRs without leaving the app.
Track every single token.
View tokens and costs by provider and model, plus your quota limits and reset times.
More capabilities
Works autonomously
- Goal execution
Give it a goal and the session works on its own until it's done.
- Schedules & webhooks
Trigger tasks on a timer or via a custom URL.
- Mobile app
Pair your iPhone or Android phone with a QR code, continue sessions end-to-end encrypted, and get notified when work finishes.
- Command line
Use the same agent in your terminal, or run one-shot tasks headlessly with mixdog exec.
Beyond code
- Browser control
Navigate and interact with real, signed-in web pages.
- Computer control
Operate native Windows and macOS applications.
- Document editing
Create and edit Word, Excel, PowerPoint, and PDF files.
- Media studio
Generate and edit images and videos, saving them to a local gallery.
Built-in tools
- Auto reasoning effort
A small local model picks the right reasoning effort for every message and tool step in about 0.1 s. On by default.
- Memory
Remembers important details and searches through past work.
- Code Graph & Code Tidy
Queries your code structure, then formats, lints, and fixes it.
- Local models
Download and run models completely locally on your GPU.
Open code. Open benchmark.
Licensed under Apache-2.0. Verdicts, verifier outputs, usage snapshots, and the exact scripts to recompute every number on this page are available in our repository.
Command line · Node.js 22.19+ or 24+
$ npm install -g mixdog
$ mixdog
Frequently asked questions
Is it really free?
Yes. The Mixdog app is free and open source under the Apache-2.0 license. You only pay for model usage directly to the providers, using your own subscriptions or API keys.
Which subscriptions and API keys work?
We support Claude account sign-in and Anthropic API keys, ChatGPT/Codex account sign-in and OpenAI API keys, Google Gemini API keys, Grok account sign-in and xAI API keys, Muse Code (Meta) account sign-in, plus OpenRouter, DeepSeek, and OpenCode Go.
What operating systems are supported?
The desktop app runs on Windows x64, macOS (Apple Silicon and Intel), and Linux (x64 and ARM64, AppImage or .deb, glibc 2.31+). On Intel Macs, auto reasoning effort and semantic memory search aren't available. The command-line version requires Node.js 22.19+ (22.x) or 24+.
How did you measure the 63% savings?
We ran all 89 Terminal-Bench 2.1 tasks five times each (445 trials per agent). Both used the exact same model, reasoning level, unmodified timeouts, and resources. We allowed zero retries and scored them with the official Harbor verifier. All logs and scripts are public.
Can I run models on my local GPU?
Yes. Using the built-in local provider, you can download and run models directly inside the app. Currently, this requires a Windows x64 machine with an NVIDIA GPU.
Why does Windows SmartScreen show a warning?
The Windows installer isn't code-signed yet. Just click 'More info' then 'Run anyway'. The Mac app is fully signed and notarized by Apple, so it opens without warnings.
Download, sign in, done.
Forget complicated config files and YAML. Follow a quick five-step tutorial and you're ready to go.