Open source · Last checked 25 Oct 2026
Open-source computer-use agents
Open-source projects let you run computer-use agents on your own infrastructure, with any model. Listed alphabetically. Check each repository for its license and activity before you depend on it.
| Project | What it is | Scope | Good for |
|---|---|---|---|
| Agent SSimular | Framework that controls a computer through its graphical interface; Python library gui-agents. Agent S3 reported 69.9% on OSWorld (Oct 2025). | Desktop | Research-grade desktop agents |
| Browser Use | Python library that lets any LLM drive a web browser. Hosted cloud also available. | Browser | Quick start on web automation |
| Cua | Framework for building, containerising and deploying desktop agents, with agent loops for Anthropic, OpenAI, UI-TARS and OmniParser. | Desktop | Sandboxed desktop agents on Windows, macOS, Linux |
| OpenClaw | General local agent platform (files, shell, browser, messaging). Reported 350,000+ GitHub stars by May 2026. Several security issues reported. | General | Personal automation; review security carefully |
| OpenCUA | Open data, tools and models for computer-use agents (NeurIPS 2025). Its 72B model reported 45.0% on OSWorld-Verified. | Models and data | Training or fine-tuning your own agent |
| Playwright MCPMicrosoft | MCP server that gives AI assistants browser control through Playwright. | Browser | Adding browser actions to an MCP-based assistant |
| Skyvern | Browser automation that combines LLMs and computer vision. | Browser | Forms and portal workflows |
| StagehandBrowserbase | Framework with four primitives (act, extract, observe, agent) on top of browser automation; TypeScript and Python. | Browser | Mixing scripted steps with AI steps |
| UI-TARSByteDance | Open models for GUI agents, such as UI-TARS-1.5-7B, which can run locally. | Model | Running a computer-use model without an API |
Sources: Cua, cua-agent on PyPI, Verdent on Agent S, OpenCUA (NeurIPS 2025), Mastra, Naoma on OpenClaw, Simon Willison.
Framework, model or platform?
- Frameworks (Browser Use, Stagehand, Skyvern, Cua, Agent S) handle the loop of screenshots and actions. You bring a model.
- Models (UI-TARS, OpenCUA) are the "brain" you can run yourself instead of calling an API.
- Infrastructure is where the agent runs. See the infrastructure directory.
Prefer a ready-made product? See computer-use agents compared.