TuriX AI has dropped what could be the most capable open-source desktop automation tool we've seen this year. The TuriX computer-use agent (CUA) lets you talk to your computer and watch it execute complex multi-step tasks—booking flights, generating charts in spreadsheets, automating web workflows—with an 80% success rate on macOS benchmarks and a respectable 64.2% on the full OSWorld benchmark that puts it in third place overall among all submitted agents.

How It Works

TuriX operates as a vision-language model (VLM) policy agent that can click, type, and navigate any application a human can use—no app-specific APIs required. The system uses a multi-model architecture with separate "brain," "actor," "memory," and optional "planner" models configured via config.json. Want to swap out the reasoning engine? Change two lines in the config file and you're running Qwen3-VL, Gemini 3 Pro, or your own local Ollama instance instead of the default TuriX models. This hot-swappable design is exactly what the open-source community has been begging for: true model portability without code changes.

The OSWorld Numbers

The benchmark results are where things get interesting. On a self-hosted OSWorld-style Mac benchmark, TuriX achieves 80%+ success rates—impressive given that zero Linux training data was used during development. For the standard OSWorld benchmark (which runs on Linux), TuriX scored 64.2% (229.88 out of 358 tasks), landing in third place on the official leaderboard. The team specifically optimized for macOS Apple Silicon but still managed a top-3 finish against agents built natively for Linux environments. That's a flex, and they know it.

Cross-Platform And Open Source

The project went multi-platform with recent updates: Windows support lives on the multi-agent-windows branch, and Linux Ubuntu compatibility shipped March 16, 2026 via the multi-agent-linux branch. All personal and research use remains completely free—no API costs if you're running local models through Ollama. The May 11 update added downloadable TuriX SuperAgent from their official web page, while an April release introduced SuperPower 3.0.0-alpha combining CLI precision with GUI automation in a single workflow for everyday office and coding tasks.

Skills And Extensibility

For power users, TuriX includes a Skills system using markdown playbooks that let you define reusable task templates. The planner sees only skill names and descriptions to select relevant ones, while the brain gets full instruction sets to execute each step. There's also native MCP (Model Context Protocol) support for hooking into Claude Desktop or any compatible agent framework. OpenClaw integration is built-in with published ClawHub skills available at clawhub.ai/Tongyu-Yan/turix-cua.

Key Takeaways

  • Third-place OSWorld benchmark finish with 64.2% success rate, despite targeting macOS optimization
  • 80%+ success on self-hosted Mac benchmarks proves real-world desktop automation capability
  • Completely open-source and free for personal/research use—no proprietary lock-in
  • Hot-swappable model architecture via config.json supports Ollama, TuriX API, OpenAI, Anthropic, or Google models
  • Cross-platform with active branches for macOS, Windows, and Linux

The Bottom Line

TuriX is what happens when hackers actually ship instead of just publishing papers. An 80% Mac success rate combined with open-source availability and model flexibility makes this the desktop automation agent to watch—and contribute to. If you've been waiting for a truly capable computer-use agent that respects your freedom (and your wallet), your wait might be over.