15 min read

Computer use agents: every way to let AI control your PC, Mac or phone

In short: a computer use agent is an AI that does not stop at writing an answer: it acts on a real device. It opens programs, edits files, runs commands, fills in web pages. In 2026 there are five ways to get one: the official desktop apps of ChatGPT and Claude (and Perplexity on Mac), the agents built into the operating system (Copilot Actions on Windows, Gemini on Android), open-source agents you host yourself such as OpenClaw, MCP servers that turn your PC into a tool of the chat you already use, and developer tools for phones. They differ in four things: which AI they work with, whether you must sit at the computer, how much setup they need, and how much they do without asking you. The comparison table below puts them side by side; the rest of the guide helps you choose.

What a computer use agent is (and what it is not)

An agent is an AI model with tools: it decides which tool to use, looks at the result and decides the next step until the job is done. That is the whole difference from a chatbot, which only produces text and leaves the work to you.

An agent does not have to work on its own for hours. If you ask Claude to find last month's invoice on your PC and email it to your accountant, and Claude searches the folders, opens the PDF and sends the email after you confirm, that is an agent at work, even though it stops as soon as the task is finished and waits for your next message. Working with you, in turns, is one way of being an agent; working alone, around the clock, is another.

Five ways an AI can reach a device

Most of the confusion comes from mixing up how the AI gets its hands on the device. These are the five routes, and most products combine two or three of them.

If you want the details of the first two routes, see computer use vs MCP: which one to use.

All the options compared

Checked against the official documentation in November 2026. These products change every month: before you pay, check the current requirements on the vendor's page.

The official assistants: ChatGPT, Claude, Perplexity, Copilot

ChatGPT and Codex

In the ChatGPT desktop app for Windows and Mac, computer use lets ChatGPT see and operate the programs on your computer, asking permission before it uses each app. From the phone, the ChatGPT app pairs with your computer through a QR code: you start a task, follow it, approve the commands it wants to run and review the result.

Watch out for: on Windows the agent works on the active desktop, so it moves your pointer and you cannot use the PC at the same time; for jobs while you are away the PC must stay unlocked and online (OpenAI suggests a virtual machine if you want to keep your own desktop free). On Mac it can also work with the screen locked. Pick it if you already use ChatGPT and the task needs real programs on screen. Setup, the phone remote and every Windows limit are in Codex computer use on Windows. To connect ChatGPT to your PC in other ways, see ChatGPT MCP on Windows.

Claude: Cowork, Claude Code and Remote Control

In the Claude desktop app for Windows and Mac, Claude can use your computer in Cowork and in Claude Code: it prefers precise tools such as connectors and the browser, and falls back to clicking on screen when nothing else works. It is a research preview for Pro and Max plans, and the desktop app must stay open and connected. Developers can continue a Claude Code session from the phone with Remote Control, while the session keeps running on the computer.

Watch out for: Dispatch, the feature for assigning tasks from the phone to the desktop, is no longer offered to new users; people who already had it can keep using it for now (if it stops connecting, see Claude Dispatch not working). Pick it if you pay for Claude and work at your desk, or you are a developer who wants to check a session from the sofa. Every way to use Claude on your PC is in control your PC with Claude, and what changed for Windows users in Claude Cowork on Windows.

Perplexity Personal Computer

The Perplexity Mac app can work across local files, native apps and the web, and is built to run continuously: Perplexity itself recommends a Mac mini left on, which then takes tasks from the phone app. It needs a paid plan (Pro, Max or Enterprise) and macOS 15 or later. It also runs on Windows 10 and 11, but there control of installed programs is not available yet: details in Perplexity Personal Computer on Windows. Pick it if you have a spare Mac and want an assistant that keeps working on its own.

Microsoft Copilot Actions

Microsoft's agent for Windows 11 works in an agent workspace: a separate session with its own Windows account and limited permissions, so it can operate apps and files in parallel with you and you can watch or take over at any time. It started as an experimental feature you switch on in Settings and rolls out in stages; at launch it excluded the European Economic Area. Pick it when it is available to you and your tasks are simple file and app chores.

Open-source agents: OpenClaw and others

OpenClaw is the most popular open-source personal agent. You install it on a computer or a server, connect one or more models with your API keys (or a local model), and talk to it from the messaging apps you already use: Telegram, WhatsApp, Slack, Discord, Signal and others. It runs as a background service, so it can work around the clock and on a schedule, and it can use MCP servers as tools.

The software is free; you pay for the model's usage. The real cost is time and responsibility: installation from the command line, keys stored on the machine, channels to configure, updates, and deciding exactly who is allowed to send it orders. Pick it if you are comfortable with the command line and want full control and any model. If you are weighing it against other tools, see OpenClaw alternatives compared. Other open-source projects worth knowing: Open Interpreter (runs code and commands locally), Agent S and UI-TARS Desktop (agents that work through the screen).

MCP servers: your PC as a tool of the chat you already use

An MCP server gives an assistant a set of tools on your computer: files, terminal, browser, programs. It does not replace your AI: it plugs into Claude, ChatGPT and the other clients that support the protocol. If the term is new, start from what an MCP server is and how it works.

The key difference is local or remote. A local server, such as Windows-MCP (free, it operates the Windows interface) or the classic version of Desktop Commander (free, terminal and files), works only with an AI app running on the same computer: you have to sit there (the two are compared in Desktop Commander vs Windows-MCP). A remote server is reachable over the internet, so it works from Claude and ChatGPT on the web and on the phone; Desktop Commander offers this through its own cloud relay. Setting up a remote server by hand means tunnels, certificates and authentication: the options are in MCP server on Windows: every option compared, and the easiest installations in install MCP on Windows.

Where MCP DeskManager fits

DeskManager is our product, so here is when it makes sense and when it does not. It is a remote MCP server for Windows: you install it, paste one address into Claude or ChatGPT, and the PC answers from any chat, including the phone apps, with no ports to open. It has 14 tools: persistent terminals, files with previous versions kept, a dedicated browser plus your real Chrome with your logins (through a free extension), programs without an API through Windows UI Automation, email, and an Android app that brings alerts, approvals and file exchange to your phone. It acts only when you ask in a chat, and anything that cannot be undone needs your confirmation.

It makes sense if you want your Windows PC from any chat, Claude or ChatGPT, also away from your desk, without managing a server. It does not if you use a Mac (the macOS version is still in development), if you want an agent that works 24/7 on its own (OpenClaw or Perplexity are built for that), or if you only work sitting at your PC with Claude Desktop: there a free local server may be enough.

Phones: the remote control and the thing being controlled

The phone plays two different roles, and the options differ a lot between them.

The phone as a remote control for your computer is the easy part: the ChatGPT app (Codex remote), the Claude app (Remote Control, and connectors to your PC, which work in the mobile app too), the Perplexity app and OpenClaw through Telegram or WhatsApp all let you start and follow work on a computer that stays at home. The four approaches are compared in control your PC from your phone with AI.

The phone as the device the AI operates is much more limited (the details, with what each route needs, are in AI that controls your Android phone):

• Android, built in: Gemini's screen automation completes tasks inside a handful of apps (ride hailing, food delivery, coffee orders) on supported phones such as the Pixel 10 and the Galaxy S26 series. It is a beta, with daily limits and few countries.
• Android, developer route: open-source MCP servers such as mobile-mcp and agents such as DroidRun drive any app through ADB, but they need USB debugging and a computer connected to the phone. Built for testing and automation, not for everyday use.
• Android, alongside a PC: the DeskManager app does not operate other apps: it receives alerts from your PC, approves the assistant's requests, exchanges files and, if you allow it, reads and sends SMS.
• iPhone: iOS does not let one app operate the others, so no consumer agent can use your iPhone apps for you; beyond Siri and Shortcuts, only developer tools for testing (Xcode, WebDriverAgent) get close. On iPhone the phone is the remote control, not the thing being controlled.

Autonomous or working with you?

Autonomy is useful for repetitive, low-risk jobs: sorting downloads, preparing a daily report, watching a page for changes. It becomes a problem when the agent touches things that matter, because nobody is there to notice a mistake. Three risks are concrete: prompt injection (an email or a web page the agent reads contains instructions that push it to do something else), keys and passwords stored on a machine that is always on and always connected, and chains of errors where one wrong step leads to the next.

An agent that works with you trades a bit of speed for control: you see each step, approve the irreversible ones, and switch it off with one click. Neither model is wrong; what matters is matching it to how sensitive the systems are (the full comparison, with the five levels of autonomy, is in autonomous AI agents). Whatever you choose:

• prefer tools that ask before deleting, sending, paying or changing accounts;
• use a separate account or a virtual machine for risky tasks on the web;
• never expose a server, a remote desktop or ADB to the internet without authentication;
• keep a way to revoke access in a second, and know where screenshots and files end up.

How to choose

• You use ChatGPT and the task needs programs on screen → ChatGPT computer use, started from the phone with Codex remote.
• You pay for Claude and work at your desk → Claude computer use in Cowork; developers → Claude Code with Remote Control.
• You want an assistant working around the clock on a spare Mac → Perplexity Personal Computer, or OpenClaw.
• You want it free and the command line does not scare you → OpenClaw, or open-source MCP servers such as Windows-MCP and Desktop Commander.
• You want your Windows PC from any chat, Claude or ChatGPT, also from the phone, acting only when you ask → a remote MCP server such as MCP DeskManager.
• You need to automate the phone itself → Gemini on supported Android phones, or ADB tools if you are a developer.
• You only need to see the screen and click yourself → remote desktop.

Frequently asked questions

What is a computer use agent?

An AI that acts on a real device instead of only writing text: it uses programs, files, commands and websites to finish a task, through the screen or through structured tools.

Can ChatGPT control my computer?

Yes, with computer use in the ChatGPT desktop app for Windows and Mac, and from the phone through Codex remote. Agent mode on the web is different: it works on a virtual computer in the cloud, not on yours.

Can Claude control my computer?

Yes, with computer use in the Claude desktop app (Cowork and Claude Code, Pro and Max plans), or through a connector to an MCP server installed on the PC, which also works from claude.ai and the phone app.

Does the computer have to stay on?

Yes, for every option that works on your own computer. With ChatGPT on Windows it must also stay unlocked.

Is there a free option?

Yes: OpenClaw (you pay only the model usage), the open-source MCP servers such as Windows-MCP and Desktop Commander, and remote desktop apps, which however need you to click yourself.

Can AI control my iPhone?

Not in the sense of using your apps for you: iOS does not allow it. You can use the iPhone to control a computer, and Siri and Shortcuts for automations inside the phone.

Is it safe?

It depends on what you authorize and on how the tool asks before acting. Choose tools with confirmations on irreversible actions, a quick way to revoke access, and no server exposed without authentication.

Further reading

• Control your PC from your phone with AI: 4 ways
• Control your PC with Claude: every way compared
• ChatGPT MCP on Windows: how to connect your PC
• Claude computer use vs MCP: which one to use
• MCP server: what it is, how it works and how to use it