// Agent profile
CUA (Computer-Use Agent) is an open-source infrastructure platform that gives AI agents the ability to see screens, click buttons, type text, and navigate desktop applications across macOS, Windows, and Linux — no per-app API integrations required. Built by a YC S25 startup, CUA provides the full stack from low-level desktop automation (Cua Driver) through sandboxed execution environments (local VMs via Lume, cloud desktops via Fleets) to evaluation tooling (Cua Bench). The platform is model-agnostic and framework-agnostic: developers bring their own agent logic and LLM while CUA handles the execution substrate. With 25K+ GitHub stars and specialist decision models (CUA-S1) on Hugging Face, it targets teams building agents that need to interact with graphical interfaces where APIs do not exist.
CUA reframes computer use for AI agents. Instead of requiring per-application API integrations, CUA gives agents the same interface a human user has — screens to look at, buttons to click, and text to type. The platform handles the full stack: a low-level automation driver that operates at the window level (not whole desktops), sandboxed execution environments for safe agent operation, and evaluation tooling for measuring agent performance on standardized tasks.
The project emerged from Y Combinator’s Spring 2025 batch and has grown to over 25,000 GitHub stars. Its approach targets what the team calls “Computer-Use 2.0” — agents that move fluidly between code, APIs, and graphical interfaces depending on which interaction mode is most effective for the task at hand.
CUA is built as a modular stack with four main components:
Cua Driver is the automation layer. It exposes click, type, scroll, screenshot, and accessibility-tree inspection across macOS, Linux, and Windows. Unlike full-desktop VNC approaches, the Driver operates on individual windows, reducing noise and improving precision. It can run as an MCP server or standalone CLI.
Lume provides local virtual machines on Apple Silicon with up to 97% native CPU speed. This gives agents isolated environments for potentially destructive operations — file system changes, application installations, system configuration — without risk to the host.
Fleets extends isolation to the cloud, provisioning managed sandboxed desktops at run.cua.ai. The Sandbox SDK is shared between local VMs and cloud Fleets, so agent code runs identically in both environments.
CUA-S1 is a family of specialist decision models trained specifically for computer-use tasks. Rather than using a general-purpose LLM for every UI interaction, CUA-S1 handles fast bounded decisions on structured interface elements like forms, dropdown menus, and dialog boxes — scoring actions rather than generating tokens.
CUA excels at automating workflows in applications that lack APIs: legacy enterprise software, CAD tools, insurance and compliance platforms with heavy form-filling, and multi-application orchestration tasks that span several desktop programs. It is also used for data collection from GUI-only sources, cross-platform testing, and training data generation for computer-use models via its benchmarking suite.
The platform is open source under the MIT license, though cloud Fleets are pay-per-use. Local VM features require Apple Silicon for macOS guests. The Driver’s window-level approach is more precise than full-desktop automation but requires the target application to be running. Some components (OmniParser, optional ultralytics) carry different licenses that may affect commercial use.
CUA is built for developers and teams building agents that need to interact with graphical interfaces — RPA automation, legacy software integration, cross-platform testing, and research into computer-use agent capabilities. It appeals to both product teams building commercial automation solutions and researchers studying agent-computer interaction.
Builder.io's open-source framework for building agent-native applications — shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together.
CLI orchestration tool that installs and configures platform connectors for 16+ social and web sources — Twitter/X, Reddit, YouTube, GitHub, Bilibili, and more. Zero API fees.
Persistent memory layer for AI coding agents — benchmark-backed (95.2% on LongMemEval-S), 92% fewer tokens per session vs full-context pasting, zero manual memory.add() calls.