Cua releases open-source computer-use agent stack and CUA-S1 models

Cua has released its open-source computer-use agent stack, providing developers with the infrastructure to give AI agents direct access to desktop environments. The project, detailed in the GitHub repository, combines isolated cloud desktops, local macOS and Linux virtual machines, and a family of specialized small models called CUA-S1. The core value proposition is enabling agents to move between code, APIs, and graphical interfaces within the same task, a concept the team labels "Computer-Use 2.0."

According to the project documentation, the stack is designed for agents that need to inspect and operate native desktop applications on macOS, Windows, and Linux. Unlike traditional automation that might rely on screen scraping or brittle UI selectors, Cua provides a Sandbox SDK that allows agents to run commands, capture screenshots, and interact with apps inside isolated environments. The release includes a "Fleet" system for maintaining sandbox capacity in the cloud, allowing code to claim a desktop from a pool and interact with it via the SDK.

The CUA-S1 Model Family

A key component of this release is CUA-S1, described as a family of small, specialized "System 1" models for computer use. The team uses the term "System 1" as an engineering analogy for fast, bounded decisions, such as choosing which value belongs in a field or whether to leave an element alone. It is explicitly not a replacement for a general-purpose agent's planning and reasoning capabilities.

The first research profile in this family, CUA-S1-FORMS, focuses on forms. Instead of generating a response token by token, the model scores decisions from structured interface elements and document values. Application code orders the actions, while the optional Cua Driver integration handles execution with explicit action boundaries. The model weights are hosted on Hugging Face, while the Python model code, synthetic-data generation, training, and evaluation scripts are available in the GitHub repository under an MIT license.

Infrastructure and Local Sandboxes

For developers who prefer local execution or need specific hardware capabilities, Cua provides tools to create and manage local macOS and Linux VMs on Apple Silicon using Apple's Virtualization.Framework. The documentation highlights a tutorial for creating a vanilla macOS Tahoe VM from an Apple restore image, starting it, and connecting over SSH using the Lume CLI. This approach allows for unattended setup defaults and provides a secure, isolated environment for agent testing.

The project also offers a benchmarking tool called Cua-Bench. Developers can build computer-use tasks, evaluate agents, and export trajectories for training. The initial setup requires Python 3.12 or 3.13 and uv installed. A basic test involves creating a small task, running its reference solution, and verifying that the evaluator reports a reward of 1.0. This benchmarking capability is crucial for reproducibility, as the project encourages users to include the specific Cua release or commit used in their experiments when citing the software.

Integration and Developer Workflow

Cua integrates with existing agent frameworks such as Claude Code, Codex, Cursor, and OpenClaw. The connection can be made through the CLI, MCP (Model Context Protocol), or typed SDKs. A notable feature is "background delivery," which allows agents to work without moving the user's pointer or taking focus, provided the app and platform support it. This is particularly useful for developers who want to run automation tasks in the background while continuing to work in the foreground.

The documentation provides a concrete example of a first result: connecting an agent, asking it to compute 6 × 7 in the Calculator app, and having it verify that the app displays 42. This simple test case demonstrates the end-to-end workflow of platform setup, permissions, and agent connection. For cloud-based workflows, the tutorial covers provisioning a Linux desktop, running uname -a, saving a screenshot, and deleting the cloud resources to avoid retaining paid capacity.

What It Means for Developers

For developers building AI agents, Cua addresses a significant gap in the current tooling landscape. Most LLMs are trained on text and code, but real-world tasks often require interacting with graphical user interfaces. Cua provides the "computer" and automation tools, allowing developers to bring their own agent and model. This modular approach means that the decision-making logic can remain with the LLM, while Cua handles the execution layer.

The specialized CUA-S1 models offer a potential efficiency gain for specific tasks like form filling. By using a small, fast model for bounded decisions, developers can reduce the latency and cost associated with using a large general-purpose model for every single UI interaction. The MIT license of the source code and the availability of model weights on Hugging Face make it accessible for both commercial and open-source projects.

However, developers should be aware of the differences between local sandboxes and cloud Fleets. While they share the Sandbox SDK, credentials, images, operations, and runtime requirements differ. The documentation advises using the runtime support reference to choose the appropriate environment. Additionally, the CUA-S1 models are described as an early, source-only research release, so users should check each model and dataset card for its scope, limitations, and artifact-specific license.

What to Watch

• Platform Support Boundaries: The documentation notes that background delivery depends on app and platform support. Developers should test their specific use cases on macOS, Windows, and Linux to understand the limitations.
• Model Maturity: CUA-S1 is a research release. While the forms model is a specific use case, the broader application of these specialized models to other UI tasks will need to be evaluated.
• Cost Management: Cloud Fleets can retain paid capacity after a claim ends. Developers should follow the cleanup steps in the tutorial to avoid unexpected costs.
• Integration Depth: As Cua integrates with major agent frameworks like Claude Code and Cursor, the depth of these integrations and the availability of community examples will be key indicators of the project's adoption.