UFO is Microsoft’s open-source multi-agent framework for automating Windows through natural language. Its latest generation, UFO³, weaves individual device agents into a coordinated “galaxy” that can operate applications across multiple machines. The project has about 10,100 stars on GitHub and is under active development in 2026.
Tell UFO what you want done — “pull this month’s numbers from Excel into a report” — and its agents plan the task, inspect the screen, and drive real applications by clicking, typing and reading UI controls. A HostAgent supervises the overall goal while per-application AppAgents specialize in software like Excel, Word, Outlook and browsers, using Windows UI Automation and vision models to ground every action in the actual interface rather than fragile scripts. UFO³ extends this from one desktop to many: a constellation layer dispatches subtasks across devices and services, so workflows can span a workstation, a laptop and cloud resources. Researchers use it as the basis of the Windows Agent Arena benchmark, and developers can plug in different LLMs as the reasoning core. UFO’s agents explain their plans before acting and keep detailed action logs, which matters when software is being driven on a user’s real machine. The framework supports hybrid grounding that combines accessibility APIs with screenshots, so it can operate both modern applications with rich UI Automation support and legacy software that exposes little structure. Microsoft Research maintains the project and publishes its benchmarks openly, inviting community-contributed agents for additional applications.
UFO is completely free and open source under the MIT license. You only pay for any model API usage your agents consume.
UFO is for researchers and developers building GUI agents, and for Windows power users who want an open platform for automating multi-application desktop workflows.