TuriX CUA

TuriX is an open-source computer-use agent built by Shenzhen-based NGT-LTD (新升代科技), designed to let AI models operate a desktop the same way a human would — clicking, typing, and navigating across any application without requiring app-specific APIs. The core repository is published under the MIT License on GitHub and also ships as a downloadable desktop app called TuriX SuperAgent. The project targets both personal productivity and enterprise automation workflows.

What It Is

TuriX CUA (Computer-Use Agent) is a Python-based agent framework that combines a vision-language model (VLM) "brain" with a specialized "actor" model to perceive screen content and execute GUI actions. It works across browsers, office suites, messaging apps, and in-house tools — anything a human can click, TuriX can attempt to automate. The agent architecture separates planning, acting, and memory into distinct model roles, each configurable via a single config.json file.

Benchmark Performance

According to the project's README, TuriX scores 64.2% (229.88 / 358) on the full OSWorld benchmark, ranking 3rd overall among all submitted agents at 50 steps. On the team's self-hosted OSWorld-style Mac benchmark, the README reports an 80%+ success rate. The README notes that TuriX was built and optimized for macOS and used zero Linux training data to achieve its OSWorld ranking.

Architecture and Key Capabilities

TuriX separates its agent into four configurable model roles — brain, actor, memory, and planner — each pointing to any compatible VLM provider:

• Hot-swappable models: Switch between TuriX API models, Gemini, Qwen3-VL, Ollama local models, or any OpenAI-compatible endpoint via config.json
• No app-specific APIs required: Operates through the GUI layer, so it works with WhatsApp, Excel, Outlook, and proprietary enterprise tools
• MCP-ready: Exposes a Model Context Protocol interface so Claude for Desktop or any MCP-compatible agent can call TuriX as a skill
• Skills system: Markdown playbooks stored in a skills/ folder; the planner selects relevant skills by name and description, while the brain receives full instructions
• Task resume: Persistent memory via memory.jsonl allows interrupted tasks to be resumed with the same agent_id

Platform and Setup Path

The open-source repository supports macOS 15+ (Apple Silicon), Windows 10+, and Linux (Ubuntu tested), each on separate branches. Setup requires Python 3.12, a conda environment, and macOS Accessibility + Safari Automation permissions. A pre-built desktop app (TuriX SuperAgent) is also available for download from the official website for users who prefer not to run from source.

Update: TuriX SuperPower 3.0.0-alpha and SuperAgent Launch

The README's latest news section records several significant milestones:

• May 11, 2026: TuriX SuperAgent became available for download from the official website
• April 8, 2026: TuriX SuperPower 3.0.0-alpha launched for macOS Apple Silicon, combining TuriX CUA + CLI in one workflow and adding TuriX-work (office task orchestration) and TuriX-code (coding and engineering automation)
• March 16, 2026: Linux support added on the multi-agent-linux branch
• Earlier 2026: v0.3 shipped DuckDuckGo integration, Ollama support, recoverable memory compression, and the Skills system; Qwen3-VL support and TuriX API model upgrades followed

The latest GitHub release is tagged v0.4, published March 6, 2026. The repository had 3,071 stars and 313 forks as of the last recorded update.

Open-Source and Commercial Dual Track

The core CUA framework is MIT-licensed and free for personal and research use. A commercial SaaS product at ngtechai.com offers a subscription-based desktop app with credit-based task execution, enterprise customization, and a dedicated TuriX API platform at turixapi.io. The open-source repo and the commercial app share the same underlying agent technology but serve different deployment needs.