Talk to your computer, watch it work.
Join our Discord community for support, discussions, and updates:
Or contact us with email: contact@turix.ai
TuriX lets your powerful AI models take real, hands‑on actions directly on your desktop. It ships with a state‑of‑the‑art computer‑use agent (achieves 80% success rate on our OSWorld‑style Mac benchmark and 64.2% success rate on OSWorld) yet stays 100 % open‑source and cost‑free for personal & research use.
Prefer your own model? Change in config.json and go.
- 📞 Contact & Community
- 🤖 OpenClaw Skill
- 📰 Latest News
- 🖼️ Demos
- ✨ Key Features
- 📊 Model Performance
- 🚀 Quick‑Start (macOS 15+)
- 🤝 Contributing
- 🗺️ Roadmap
Use TuriX via OpenClaw with our published ClawHub skill:
https://clawhub.ai/Tongyu-Yan/turix-cua
This repo also includes local OpenClaw skill packages in OpenCLaw_TuriX_skill/:
- macOS package in
main(SKILL.md+scripts/run_turix.sh) - Windows package in
multi-agent-windows(SKILL.md+scripts/run_turix.ps1+agents/openai.yaml)
For installation and permissions, follow OpenCLaw_TuriX_skill/README.md.
April 8, 2026 - 🚀 Introducing TuriX SuperPower 3.0.0-alpha for macOS (Apple Silicon):
Download dmg(ONLY for mac)
This is our all-in-one productivity app that combines TuriX CUA + CLI in one workflow, and adds two new capabilities:
- TuriX-work for everyday office execution and task orchestration
- TuriX-code for coding, automation, and engineering tasks
From writing code to handling office tasks, you can execute with CLI precision and close the loop through GUI actions in one continuous flow.
March 16, 2026 - 🐧 Linux support is now available on branch multi-agent-linux. If you want to run TuriX on Linux (for example Ubuntu), switch to that branch first:
git checkout multi-agent-linuxMarch 9, 2026 - Added a new OpenClaw Flash/Fast Mode skill for macOS on branch mac_legacy. If you want to use this faster, lighter setup, switch to that branch first:
git checkout mac_legacyMarch 5, 2026 - Updated the Windows OpenClaw local skill on branch multi-agent-windows with direct dispatch, safer pre-flight checks, and the new OpenCLaw_TuriX_skill/agents/openai.yaml.
Earlier updates (Jan 2026 and before) - We shipped v0.3 (DuckDuckGo, Ollama, recoverable memory compression, Skills), published the TuriX OpenClaw skill on ClawHub, upgraded the core architecture to multi-model, and rolled out major model capability improvements including Qwen3-VL support and TuriX API model upgrades.
Ready to level up? Update your config.json and start automating—happy hacking! 🎉
Stay tuned to our Discord for tips, user stories, and the next big drop.
TuriX SuperPower App Demo
Book a flight, hotel and uber.
Search iPhone price, create Pages document, and send to contact
Generate a bar-chart in the numbers file sent by boss in discord and insert it to the right place of my powerpoint, and reply my boss.
Search video content in youtube and like it
Claude search for AI news, and call TuriX with MCP, write down the research result to a pages document and send it to contact
| Capability | What it means |
|---|---|
| SOTA default model | Outperforms previous open‑source agents (e.g. UI‑TARS) on success rate and speed on Mac |
| No app‑specific APIs | If a human can click it, TuriX can too—WhatsApp, Excel, Outlook, in‑house tools… |
| Hot‑swappable "brains" | Replace the VLM policy without touching code (config.json) |
| MCP‑ready | Hook up Claude for Desktop or any agent via the Model Context Protocol (MCP) |
| Skills (markdown playbooks) | Planner selects relevant skill guides (name + description), brain uses full instructions to plan each step |
Our agent achieves state-of-the-art performance on desktop automation tasks:
TuriX scores 64.2% (229.88 / 358) on the full OSWorld benchmark, ranking 3rd overall among all submitted agents. Notably, TuriX is built and optimized for macOS, where we achieve an 80%+ success rate on our self-hosted OSWorld-style Mac benchmark. We used zero Linux training data, yet still achieve a top-3 finish on OSWorld's Linux-based environment.
For more details, check our report.
We never collect data—install, grant permissions, and hack away.
0. Windows Users: Switch to the
multi-agent-windowsbranch for Windows-specific setup and installation instructions.git checkout multi-agent-windowsFor the updated OpenClaw Windows local skill package, see
OpenCLaw_TuriX_skill/README.mdin that branch.0. Linux Users: Switch to the
multi-agent-linuxbranch for Linux-specific setup and installation instructions.git checkout multi-agent-linux0. Windows Legacy Users: For the previous Windows setup, switch to the
windows_legacybranch.0. macOS Legacy Users: For the previous single-model macOS setup, switch to the
mac_legacybranch.
For easier usage, download the app
Or follow the manual setup below:
Firstly Clone the repository and run:
conda create -n turix_env python=3.12
conda activate turix_env # requires conda ≥ 22.9
pip install -r requirements.txt- Open System Settings ▸ Privacy & Security ▸ Accessibility
- Click ＋, then add Terminal and Visual Studio Code ANY IDE you use
- If the agent still fails, also add /usr/bin/python3
- Safari ▸ Settings ▸ Advanced → enable Show features for web developers
- In the new Develop menu, enable
- Allow Remote Automation
- Allow JavaScript from Apple Events
# macOS Terminal
osascript -e 'tell application "Safari" \
to do JavaScript "alert(\"Triggering accessibility request\")" in document 1'
# VS Code integrated terminal (repeat to grant VS Code)
osascript -e 'tell application "Safari" \
to do JavaScript "alert(\"Triggering accessibility request\")" in document 1'
Click "Allow" on every dialog so the agent can drive Safari.
Important
Task Configuration is Critical: The quality of your task instructions directly impacts success rate. Clear, specific prompts lead to better automation results.
Edit task in examples/config.json:
{
"agent": {
"task": "open system settings, switch to Dark Mode"
}
}Get API now with $20 credit from our official web page. Login to our website and the key is at the bottom.
In this main (multi-agent) branch, you need to set the brain, actor, and memory models. It only supports mac for now. If you enable planning
(agent.use_plan: true), you also need to set the planner model.
We strongly recommand you to set the turix-actor model as the actor. The brain can be any VLMs you like, we provide qwen3.5vl in our platform. Gemini-3-pro is tested to be smartest, and Gemini-3-flash is fast and smart enough for most of the tasks.
Edit API in examples/config.json:
"brain_llm": {
"provider": "turix",
"model_name": "turix-brain",
"api_key": "YOUR_API_KEY",
"base_url": "https://turixapi.io/v1"
},
"actor_llm": {
"provider": "turix",
"model_name": "turix-actor",
"api_key": "YOUR_API_KEY",
"base_url": "https://turixapi.io/v1"
},
"memory_llm": {
"provider": "turix",
"model_name": "turix-brain",
"api_key": "YOUR_API_KEY",
"base_url": "https://turixapi.io/v1"
},
"planner_llm": {
"provider": "turix",
"model_name": "turix-brain",
"api_key": "YOUR_API_KEY",
"base_url": "https://turixapi.io/v1"
}For a local Ollama setup, point each role to your Ollama server:
"brain_llm": {
"provider": "ollama",
"model_name": "llama3.2-vision",
"base_url": "http://localhost:11434"
},
"actor_llm": {
"provider": "ollama",
"model_name": "llama3.2-vision",
"base_url": "http://localhost:11434"
},
"memory_llm": {
"provider": "ollama",
"model_name": "llama3.2-vision",
"base_url": "http://localhost:11434"
},
"planner_llm": {
"provider": "ollama",
"model_name": "llama3.2-vision",
"base_url": "http://localhost:11434"
}If you want to use other models not defined by the build_llm function in the main.py, you need to first define it, then setup the config.
main.py:
if provider == "name_you_want":
return ChatOpenAI(
model="gpt-4.1-mini", api_key=api_key, temperature=0.3
)
Switch between ChatOpenAI, ChatGoogleGenerativeAI, ChatAnthropic, or ChatOllama base on your llm. Also change the model name.
Skills are lightweight markdown playbooks stored in a single folder (default: skills/). Each skill file starts with YAML frontmatter containing name and description, followed by the instructions. The planner only sees the name + description to select relevant skills; the brain receives the full skill content to guide step goals.
Skills selection requires planning (agent.use_plan: true).
Example skill file (skills/github-web-actions.md):
---
name: github-web-actions
description: Use when navigating GitHub in a browser (searching repos, starring, etc.).
---
# GitHub Web Actions
- Open GitHub, use the site search, and navigate to the repo page.
- If login is required, ask the user before proceeding.
- Confirm the Star button state before moving on.Enable in examples/config.json:
{
"agent": {
"use_plan": true,
"use_skills": true,
"skills_dir": "skills",
"skills_max_chars": 4000
}
}python examples/main.pyEnjoy hands‑free computing 🎉
To resume a task after an interruption, set a stable agent_id and enable resume in examples/config.json:
{
"agent": {
"resume": true,
"agent_id": "my-task-001"
}
}Notes:
- Use the same
agent_idas the run you want to resume. - Keep the same
taskwhen resuming. - Resume only works if prior memory exists at
src/agent/temp_files/<agent_id>/memory.jsonl. - To start fresh, set
resumetofalse, changeagent_id, or deletesrc/agent/temp_files/<agent_id>.
We welcome contributions! Please read our Contributing Guide to get started.
Quick links:
For bug reports and feature requests, please open an issue.
| Quarter | Feature | Description |
|---|---|---|
| 2025 Q3 | ✅ Terminate and Resume | Support resuming from terminated task. |
| 2025 Q3 | ✅ Windows Support | Cross-platform compatibility bringing TuriX automation to Windows environments (Now Available) |
| 2025 Q3 | ✅ Enhanced MCP Integration | Deeper Model Context Protocol support for seamless third-party agent connectivity (Now Available) |
| 2025 Q4 | ✅ Next-Gen AI Model | Significantly improved clicking accuracy and task execution capabilities |
| 2025 Q4 | ✅ Windows-Optimized Model | Native Windows model architecture for superior performance on Microsoft platforms |
| 2025 Q4 | ✅ Support Gemini-3-pro model | Run with any compatible vision language models |
| 2025 Q4 | ✅ Planner | Understands user intent and makes step-by-step plans to complete tasks |
| 2025 Q4 | ✅ Multi-Agent Architecture | Evaluate and guide each step in working |
| 2025 Q4 | ✅ Duckduckgo Integration | Speed up the information gathering process, for smarter planning (now on main) |
| 2026 Q1 | ✅ Ollama Support | Support the Ollama Qwen3vl models |
| 2026 Q1 | ✅ Recoverable Memory Compression | Advance memory management mechanism to stabelize performance (Commited beta version) |
| 2026 Q1 | ✅ Skills | Stablize the agent workflow. |
| 2026 Q1 | ✅ OpenClaw Skill | Published on ClawHub (https://clawhub.ai/Tongyu-Yan/turix-cua) so OpenClaw can use TuriX as its eyes and hands. |
| 2026 Q1 | ✅ OpenClaw Windows Skill Refresh | Updated multi-agent-windows local skill package with direct dispatch (turix/turix-win), required-branch checks, and --dry-run support. |
| 2026 Q1 | ✅ Linux Support | Linux support is now available on branch multi-agent-linux (Ubuntu and other distributions). |
| 2026 Q2 | Browser Automation | Support a Chrome-like browser for scalability |
| 2026 Q2 | Persistent Memory | Learn user preferences and maintain task history across sessions |
| 2026 Q2 | Learning by Demonstration | Train the agent by showing it your preferred methods and workflows |