Cua Spaces wants to enable AI agents to work on the computers of their choice

Cua introduces Spaces, a SDK to run Claude Code, Codex, and other AI agents on your own computers, locally or in the cloud.

Computer-use agents are typically tied to a execution environment chosen by their provider. With Cua Spaces, Cua offers to decouple these two elements: developers choose their AI agent, as well as the machine it works on. An assistant like Claude Code or Codex can thus access a graphical desktop, open applications, run commands, and carry out a task on a local computer, a remote Mac, or cloud infrastructure. The new SDK comes with an application that lets users monitor operations live, take over control, and distribute multiple agents across different workspaces. A Dedicated Desktop for Each Agent

Cua Spaces relies on a simple idea: providing AI agents with a computing environment they can use like a regular computer. Each Space corresponds to a desktop with its own screen, applications, and resources. The agent can run commands, manipulate files, or interact with a graphical interface, while the user watches its work unfold.

The main benefit of this approach lies in separating the agent from the hardware. Unlike solutions that impose a virtual machine hosted by the model provider, Cua leaves the choice of infrastructure to the user. A Space can run on the user's computer, on another machine in their team, or in a cloud environment they manage. The platform specifically mentions AWS, Google Cloud, Azure, and isolated environment providers like E2B, Daytona, or Modal.

This flexibility allows developers to maintain control over specific hardware and software configurations: operating system, available resources, installed applications, and data location. This does not mean agents automatically run locally. Claude Code, Codex, or other assistants can continue to use models hosted on remote servers, even when their workspace is located on a personal machine. Claude Code, Codex, and Other Agents in the Same Environment

Cua is not just building another agent. Its SDK aims to provide a shared execution layer for multiple existing assistants. Compatible environments include Claude Code, Codex, Cursor, Gemini CLI, Hermes, OpenClaw, and OpenCode. Developers can also build their own integrations using the provided tools.

The connection is established notably through the Model Context Protocol (MCP), which exposes computing capabilities to an assistant. Cua provides an MCP server equipped with skills and tools to discover available computers, create Spaces, run commands, transfer files, or view a machine's screen. Agents thus use a common interface to access different environments.

Installation can be done via the app or through a terminal. A command like `cua agents setup --agents claude,codex --yes` configures the corresponding integrations. Users can then ask their assistant to list available Spaces or create an environment for a specific task.

The entire setup is designed to reduce the friction required when a team uses multiple agents. Developers can keep their preferred tools while choosing where operations take place, without having to rebuild a separate infrastructure for each assistant. Distributing Tasks Across Multiple Computers

Beyond creating a virtual desktop, Cua Spaces makes it possible to manage multiple environments from a single interface. The SDK offers functions to start an agent, send it an instruction, check its status, and stop its execution. Sessions can continue running inside the Space even when the client app is disconnected.

This architecture enables task distribution across multiple machines. For instance, a team can dedicate one environment to running software tests, a second to analyzing files, and a third to browser-based tasks. The Spaces remain distinct, making it easy to allocate the necessary applications and resources for each activity.

In the TypeScript example published by Cua, a developer starts a Claude Code agent inside a Space, instructs it to open a repository and run a test suite, and then sends a follow-up instruction to fix any failures. The program can then check the agent's status and close the session, with the option to automatically delete the environment used.

The platform also supports shared usage modes. Multiple operations can run within the same Space, though Cua warns that isolation between agents is no longer guaranteed in this case. When tasks must remain independent or handle different data, a dedicated environment remains preferable. Monitoring Work and Taking Control

Cua Spaces expands interaction beyond the terminal. The desktop app displays available environments and lets users watch their screen live, even when operations run remotely. Users can interact with the interface, take control of the desktop, or move files and applications between certain environments.

This visibility addresses one of the practical hurdles of computer-use agents: understanding what they are doing during multi-step tasks. An assistant might open a browser, adjust settings, or interact with software without the final result alone explaining all the steps taken. Visual tracking allows users to verify progress and intervene before execution finishes.

Cua also includes features for collaborative work. A Space can be shared with a colleague in read-only mode, following user confirmation. This feature can be used to showcase task progress, inspect a test environment, or observe an agent's behavior without automatically granting permission to modify the machine. Persistent Agents Capable of Continuing Their Tasks

The SDK supports long-running agent sessions structured as execution threads. An agent can be started in an environment, receive new instructions, and maintain its working context across multiple interactions. Management functions allow users to view active runs, retrieve outputs, or request an execution stop.

Cua also offers persistent agents equipped with a workspace, memory, and configurable routines. These agents can be linked to specific authorized machines based on user permissions. The goal is to allow them to resume recurring tasks without rebuilding their