On this page
- What does a computer-use agent need from a Mac?
- How do Claude computer use and OpenAI computer use reach a Mac?
- Which options run computer-use agents on macOS?
- Revyl
- Cua
- E2B
- Tart
- Raw Macs: MacStadium, MacinCloud and AWS EC2 Mac
- How do agents connect to a Mac sandbox?
- What does each option do best?
- Beyond macOS computer use: what else Revyl runs
- Sources
- Frequently asked questions
A computer use agent needs a Mac it can see and control. A Revyl Mac computer starts on demand and gives the agent its desktop, shell and files; the same CLI opens web browsers, iOS simulators and Android emulators. Revyl's Free plan includes up to 1,200 free device minutes a month on iOS, Android, web and Mac with no card and no monthly fee; past that, a Revyl Mac computer is $0.05 a minute.
Scroll the table sideways to see all 3 columns.
A computer use agent gets a Mac in one of three ways: it starts a hosted Mac computer on demand from Revyl, it works in a macOS sandbox or desktop that you run on your own Apple hardware with Tart or Cua, or it connects to a whole Mac rented from MacStadium, MacinCloud or AWS EC2 Mac. E2B is hosted too, and its computer-use desktop runs Linux.
What does a computer-use agent need from a Mac?
Computer use means the agent reads the screen and acts on it. It takes a screenshot, decides what to do and sends clicks, keystrokes and scrolls, so its Mac needs a desktop the agent can capture and drive. Most macOS work also goes faster with a shell, because running xcodebuild or git beats clicking through menus. And a person usually wants to watch the desktop and see afterwards what the agent did.
How do Claude computer use and OpenAI computer use reach a Mac?
Both are model tools, and neither ships the computer. Claude computer use gives the model screenshot, mouse and keyboard tools, and your application runs every call in an environment you control. OpenAI computer use connects the API loop to a browser or desktop runtime you supply. For macOS, that runtime is a Mac the harness can screenshot and send input to.
With a Revyl Mac computer, the harness passes each revyl computer screenshot to the model and runs the actions it returns with click, type, key, scroll and drag (CLI reference). Chat apps such as Claude and ChatGPT can also connect to Revyl’s hosted MCP server, which starts a macOS computer, takes a screenshot and stops it.
Which options run computer-use agents on macOS?
Revyl
Revyl Mac computers are macOS computers in the cloud that an agent or a person starts, uses and stops. An agent starts one from the CLI or the MCP server, runs programs such as git or xcodebuild for up to an hour per command, reads and writes files, and drives the desktop with screenshot, click, type and key commands, where a click can name its target, such as “the Save button” (CLI reference). People watch the desktop live in the browser and can take control. Web and iOS apps run inside the computer as app sessions with their own reports, and GitHub Actions jobs run on a Revyl Mac with runs-on: revyl-macos. The computer’s page keeps an Activity log of every command, file and process action. A computer runs for up to two hours from creation (Mac computers), and Mac computers are $0.05 a minute after the free monthly allowance (pricing).
Cua
Cua is open-source infrastructure for computer-use agents. Cua Spaces shares desktops on computers you own or rent, Lume runs up to two macOS VMs at a time on an Apple silicon Mac, and a macOS Space with developer tools gives agents Xcode and the iOS Simulator. Cua’s hosted Cua Cloud ended on 18 October 2026, per its migration guide. See Revyl vs Cua.
E2B
E2B runs each agent session in an isolated Firecracker microVM. Its desktop for computer use runs Linux, with a Desktop SDK for mouse, keyboard and screen capture, and sandboxes run up to 24 hours continuously on Pro, per its lifetime docs. See Revyl vs E2B.
Tart
Tart runs macOS and Linux VMs on Apple Silicon Macs you own, with VM images in OCI-compatible registries and Orchard for clusters. Teams use it to build their own macOS sandbox, then add the agent’s screen and input tooling on top.
Raw Macs: MacStadium, MacinCloud and AWS EC2 Mac
- MacStadium rents dedicated Macs and runs Orka macOS VMs.
- MacinCloud rents managed Mac accounts and servers reached over RDP or a browser.
- AWS EC2 Mac runs one macOS instance per Dedicated Host in your AWS account.
On a raw Mac, the agent connects the way a person does, over SSH, RDP or Apple Remote Desktop, and you set up its screen capture and input yourself.
How do agents connect to a Mac sandbox?
However you set up macOS for AI agents, each agent needs a way to start the Mac, see its screen and run commands on it. Cua connects agents through Cua skills and its MCP server, and E2B through its SDKs. An agent starts, uses and stops a Revyl Mac computer itself through the Revyl CLI or MCP server, drives the desktop with screen commands, and leaves an Activity log of each command, file and process action. For how agent sandboxes compare beyond macOS, see agent sandboxes explained.
What does each option do best?
- Revyl: a hosted Mac that a computer use agent starts, drives and stops on demand, with a shell, files and CI jobs and nothing to provision (Mac computers).
- Cua and Tart: computer-use agents on Apple Silicon hardware you own (Cua, Tart).
- E2B: a Linux desktop for computer use (E2B computer use).
- MacStadium, MacinCloud and AWS EC2 Mac: a long-lived machine you administer (MacStadium, MacinCloud, AWS EC2 Mac).
Beyond macOS computer use: what else Revyl runs
The agent that uses a Revyl Mac can also open the app it builds there on Revyl’s other runtimes.
- Web: web apps run in live Chrome sessions streamed to the browser, at a desktop or mobile viewport, on every plan (web sessions).
- iOS simulators: cloud iOS simulator sessions start from the dashboard, the CLI or an agent (sessions).
- Android emulators: cloud Android emulators stream live to the browser in the same way (sessions).
- macOS computers: a hosted Mac computer at $0.05 a minute, with simulator and emulator runs at $0.075 per device minute on iOS and Android (pricing).
- Atlas: Atlas is the shared record of what the app does, filled in from the screens, states and transitions observed in Revyl tests, sessions and explorations.
- Agent access: agents start and drive every runtime through the Revyl CLI and MCP server.
- Reports and PR verdicts: each test run produces a report with video, a step timeline, performance metrics, device logs and network traffic (on iOS and debuggable Android builds). With proof of changes, Revyl runs the changed flows on each pull request’s build and posts a verdict.
Sources
Facts were checked against these public pages on November 11, 2026. Spot an error? Email support@revyl.ai and we will correct it.
- Claude computer use tool
- OpenAI computer use guide
- Cua Spaces pricing
- Lume
- Cua macOS Space with developer tools
- E2B enterprise
- E2B computer use
- E2B sandbox lifetime
- Tart
- MacStadium pricing
- Orka overview
- MacinCloud Managed Server
- Amazon EC2 Mac FAQs
- EC2 Mac user guide
- Revyl sessions
- Revyl Mac computers
- Revyl GitHub Actions on Mac computers
- Revyl CLI reference
- Revyl MCP setup
- Revyl reports
- Revyl Atlas
- Revyl proof of changes
- cua.ai/docs/spaces/guides/connect-your-agent
- e2b.dev/security
- tart.run/quick-start
- support.macincloud.com/support/solutions/articles/8000051073-access-macincloud-managed-servers-through-rdp-client
- aws.amazon.com/ec2/instance-types/mac
Frequently asked questions
What is a computer use agent?
A computer use agent is an AI agent that operates software through its screen. It takes a screenshot, decides what to do and sends clicks, keystrokes and scrolls. On macOS it needs a Mac whose screen it can read and whose input it can drive, which is what a Revyl Mac computer, a Cua desktop or a macOS sandbox on your own Mac provides.
Can Claude computer use or OpenAI computer use work on a Mac?
Yes. Both are model tools: the model returns actions, and your application runs them in a browser or desktop you supply. Any Mac the harness can screenshot and send input to works. A Revyl Mac computer takes those actions through revyl computer screenshot, click, type and key, and chat apps such as Claude and ChatGPT can start one through Revyl's hosted MCP server.
What is the best way to give an AI agent a Mac?
A hosted Mac computer needs nothing provisioned: a Revyl Mac computer starts on demand from the CLI or MCP server, gives the agent a desktop, shell, files and processes, and keeps an Activity log of what it did. Tart or Cua's Lume run macOS VMs on hardware you own, and MacStadium, MacinCloud or AWS EC2 Mac rent a whole Mac. Sources: tart.run, cua.ai/docs/lume, macstadium.com/pricing, macincloud.com and aws.amazon.com/ec2/instance-types/mac.
Can a computer-use agent run the iOS Simulator?
Yes. Cua's macOS Space with developer tools includes Xcode and the iOS Simulator on your own Apple silicon Mac. Revyl starts a cloud iOS simulator session directly from its CLI or MCP server, or runs an iOS app as an app session inside a Revyl Mac computer.
Which macOS sandbox options run on my own hardware?
Tart and Cua run on Macs you own, and MacStadium's Orka can also run on-premises. E2B offers BYOC in your own cloud account on Enterprise plans, with a Linux desktop for computer use.
Can an agent control a Revyl Mac computer's screen?
Yes. The Revyl CLI takes screenshots of a Mac computer's desktop and clicks, drags, scrolls, types and presses keys on it, and a click can name its target, such as "the Save button". The same CLI runs commands, reads and writes files and lists processes.
What else does Revyl do?
Besides macOS computers, Revyl runs web apps in live Chrome sessions and mobile apps on cloud iOS simulators and Android emulators, all from the dashboard, the CLI or the MCP server. Atlas is the shared record of what the app does, filled in from observed runs, and Revyl posts a verdict on each pull request with proof of changes.