Cua

Purpose: Scaling infrastructure to train, evaluate, and generate data for computer-use AI agents across multiple operating systems.

Inferred · not functionally tested

Audience: AI researchers and developers

Inferred · not functionally tested

Functions: agents, monitoring, data_extraction

Inferred · not functionally tested

Interfaces: API: unknown · MCP: indicated (inferred, not tested) · CLI: indicated (inferred, not tested) · Self-hosting: indicated (inferred, not tested)

Recorded constraints: pricing: open_source · license: MIT · platforms: CLI · deployment: cli, self_hosted

Constraint provenance is unknown; confirm requirements with the publisher.

Record sources: cua.ai · github.com. These links do not verify the individual claims.

Cua sits in PulseGate's Frameworks & runtimes category. Inferred · not functionally tested: It focuses on scaling infrastructure to train, evaluate, and generate data for computer-use AI agents across multiple operating systems. Inferred · not functionally tested: It is built as an open-source project for AI researchers and developers. Basis unknown · not verified: The project is open source (MIT). Basis unknown · not verified: Cua is available on the command line, and it can be self-hosted.

Cua builds and maintains Cua, and it first shipped in 2025. The project is developed in the open on GitHub with 20.6k stars and 777 commits in the last 90 days. Inferred · not functionally tested: Among its 5 catalogued features are Agent Fleets, Cross-OS Support, and Background Drivers. Basis unknown · not verified: Catalogued interfaces include an MCP server.

Summary written by a language model from the project’s public pages.

Do you need to control an existing desktop or provision an isolated computer?

Cua distinguishes Driver, which controls native desktop applications, from Sandbox and Fleets, which provide environments for computer-use workloads. Its current site lists Linux, Windows, macOS and Android fleet options; availability depends on the selected environment.

Choose the execution boundary before choosing a model. A workflow in an existing user session has different access and reset requirements from a disposable environment. Treat the fleet, operating system and driver as separate choices, not interchangeable product names.

Suggested evaluation: run a harmless task in a test environment, record its initial state, then verify how you reset it and inspect a failed run. Confirm the exact operating-system backend and account requirements in the current documentation. We have not benchmarked execution reliability.

PulseGate evidence note · sources reviewed 22 Sep 2026. Document review, not a hands-on software test. Project details may change.

Sources: Cua product architecture · Cloud fleet tutorial