CUA

You can delegate desktop automation tasks to an autonomous agent that runs inside cloud-provisioned sandboxes. This MCP server handles task delegation, screen description, and progressive task execution while ensuring images stay in the sandbox and only text summaries are returned. It enables creating and managing sandboxes, describing current screens, and running complex workflows end-to-end.

How to use

Connect to the MCP server from your MCP client or integration. You can run tasks that automate desktop actions inside a sandbox, request a text-based description of the current screen, and poll task progress to see what the agent is doing. Typical flows include starting a sandbox, issuing a task such as opening an application and navigating to a URL, and then querying the screen state to verify results. All actions are performed inside the sandbox and the final output is a concise summary of what was achieved.

How to install

Prerequisites: Node.js and npm installed on your system. A modern browser or terminal access for verification is helpful. You also need access to a CUA API key and an Anthropic API key for vision processing if you plan to run images through the vision model.

Step-by-step setup to use an MCP client with this server setup:

1. Obtain a CUA API key from your account and an Anthropic API key for vision processing.

2. Add your MCP server URL to your client configuration as shown in the example below.

Configuration and usage notes

Quick start shows how to configure your client to point at the MCP server for remote operation. You can also run the MCP client locally via standard package tooling if you are integrating into an automated workflow.

Key concepts include: starting and listing sandboxes, describing the current screen state with vision, and running autonomous tasks that report back a task_id and a summary.

Troubleshooting and tips

If you encounter a missing API key error, set CUA_API_KEY in your environment or pass it via the X-CUA-API-Key header in requests.

If the vision component prompt returns an error, ensure your ANTHROPIC_API_KEY is configured and that you have network access to the vision service.

Available tools

list_sandboxes

List all CUA cloud sandboxes with their current status

get_sandbox

Get details of a specific sandbox including API URLs

start_sandbox

Start a stopped sandbox

stop_sandbox

Stop a running sandbox

restart_sandbox

Restart a sandbox

describe_screen

Get a text description of current screen state using vision AI. No actions taken.

run_task

Execute a computer task autonomously and return a task_id for polling

get_task_progress

Poll progress of a running task, including current step and reasoning

get_task_history

Retrieve results of a previously executed task by ID