DeskManager

9 min read

Claude Computer Use vs MCP in 2026: How They Work, How to Turn Them On and Which to Choose

Guide · updated November 2026 · Protime

In short: an AI can work on a computer in two ways. Computer use acts like a person: it looks at a screenshot, decides where to click, moves the mouse and keyboard, looks again. It works with any program, but every step costs seconds and tokens. MCP (Model Context Protocol) instead gives the assistant tools with a name and parameters: “run this command”, “read this file”, “press the Save button”. It is fast and precise, but can only do what the tools offer. In 2026 Claude uses both and chooses on its own: tools first, the screen as a last resort. Below: how each one works, how to turn it on step by step, and when one beats the other.

The differences in one table

AspectComputer useMCP
How it actsScreenshot, then click, type, scrollCalls tools with precise parameters and gets data back
What it can useAny program visible on screenOnly what the server's tools expose
SpeedA few seconds per stepOne call per action, often under a second
ReliabilitySensitive to pop-ups, resolution, changing layoutsHigh: results are text and data, not pixels
Token costHigh: every step sends an imageLow: only text and data go in and out
Where you turn it onClaude desktop app (Cowork and Claude Code) or the developer APIDesktop app extensions, connectors in every chat, Claude Code, API
Claude planPro, Max, Ultra (research preview); pay-as-you-go via APIAll, Free included (one custom connector)
From your phoneNo, it needs the desktop app open on the PCYes, with a remote server added as a connector
Typical usePrograms with no integrations, legacy interfacesFiles, terminal, web, apps with an API or accessibility

How computer use works

Computer use is a loop. Claude receives a screenshot, analyzes it, picks an action (click at given coordinates, type text, press a key combination, scroll, zoom into an area to read it better), performs it and asks for a new screenshot to see the result. It repeats until the task is done or it needs you.

The upside is universality: if a person can do it by looking at the screen, in principle Claude can too, even in a twenty-year-old inventory program with no integrations at all. The price is that every step is slow and fragile: an update window that pops up, a moved button or a different resolution can throw it off. Anthropic's own developer documentation recommends MCP or native integrations when they exist, and keeping the screen for what has no other way in.

In the Claude desktop app

Since 2026 computer use lives inside the desktop app on Windows and Mac, in Cowork and Claude Code, as a research preview for Pro, Max and Ultra. On Windows it needs an x64 PC (not ARM). Before touching the screen Claude tries the tools it already has: a connector, the terminal, Claude in Chrome. The screen comes in only when nothing else works.

  1. Install or update the desktop app from claude.com/download and sign in.
  2. Open Settings and, in the Desktop app section, go to General (in newer versions the option may be under Capabilities or System → This computer).
  3. Turn on Enable computer use and confirm.
  4. Start a task in Cowork or Claude Code. The first time Claude wants to use an app it asks your permission, app by app.
  5. Optionally, list in Denied apps the programs it must never open.

During the task the PC must stay on, unlocked and with the app open. Sensitive apps, such as banking, are best excluded.

Through the API, for developers

If you build your own agent, you use the computer use tool of the Claude API. In the current version the tool is called computer_20251124 and requires the beta header computer-use-2025-11-24; it supports screenshots, clicks, typing, keys, scrolling and a zoom action for reading small details. Older models use the computer_20250124 version.

  1. Create an API key in the Claude Console.
  2. Prepare an isolated environment: a virtual machine or a container with a desktop, never your everyday PC. Anthropic provides a Docker reference implementation.
  3. In the API request, declare the computer use tool with the screen resolution and add the beta header.
  4. Write the loop: when Claude asks for an action, perform it in the environment (with a tool such as xdotool or pyautogui), take the screenshot and send it back as the result.
  5. Set limits: a maximum number of steps, and human confirmation for actions that spend money or delete data.

How MCP works

MCP is an open protocol: a server declares a list of tools, each with a name, a description and the parameters it accepts; the client (Claude, ChatGPT, another assistant) shows them to the model, which decides which to call and with what values. The server runs it and answers with data. “Find September's PDFs” becomes a file search with a date filter, not a series of clicks in File Explorer.

An MCP server can be local (it runs on the PC and talks only to the desktop app on that PC, over stdio) or remote (reachable from the internet over HTTPS, with the Streamable HTTP transport and usually OAuth sign-in). Only a remote server works from claude.ai, the phone apps and ChatGPT. The protocol is explained in full in the complete guide to MCP servers.

Adding a local MCP server (extension)

  1. In the Claude desktop app open Settings → Extensions.
  2. Pick an extension from the directory and press Install, or double-click a downloaded .mcpb file.
  3. Fill in the requested fields (for example the folder to give access to) and enable it.

Adding a remote MCP server (connector)

  1. On claude.ai or in the desktop app open Customize → Connectors.
  2. Press + and choose Add custom connector.
  3. Enter a name, paste the server's HTTPS address and, if the server requires it, open the advanced settings for the OAuth client ID.
  4. Press Add, then Connect and complete the sign-in on the page that opens.
  5. In a chat open the + menu, go to Connectors and turn on the new one. From then on it also works in the phone app.

On the Free plan you can add one custom connector; on Team and Enterprise an Owner adds it for the whole organization.

The same task done both ways

Take a simple request: “Find October's PDF invoices in the Documents folder and tell me the total.”

StepWith computer useWith MCP
Finding the filesOpens File Explorer, goes to Documents, sorts by date, scrolls: a dozen screenshotsOne file search filtered by extension and date
Reading the amountsOpens each PDF, zooms into the total, reads it from the imageReads each PDF's text and finds the total
Adding upIn its head or by opening CalculatorIn the reply or with a command
Rough timeMinutesSeconds

Now a different task: “In the old warehouse program, which has no API, enter these three items.” Here an MCP server that can only read files and run commands is no help: computer use wins, or the hybrid approach below.

The hybrid approach: the interface read by name

Between screenshots and APIs there is a third way. Windows exposes the elements of every window through its accessibility layer (UI Automation): buttons, fields, menus and tables have a name, a type and a state. An MCP tool can read that tree and say “press the Save button” or “type 12 in the Quantity field” without guessing pixels, falling back to a screenshot only when a program exposes nothing. It is faster than pure computer use and covers most desktop programs.

This is how MCP DeskManager handles Windows programs, alongside its tools for the terminal, files, browser and email. Because it is a remote server, you use it from any Claude chat, phone included, and from ChatGPT on the web.

Security: what changes between the two

How to choose

The full picture of the ways to work on your PC with Claude, with setup for each, is in control your PC with Claude.

Frequently asked questions

What is the difference between computer use and MCP?

Computer use acts on the screen with screenshots, mouse and keyboard and works with any program, but it is slow. MCP gives the assistant structured tools, fast and precise, limited to what the server exposes.

Does MCP replace computer use?

No. MCP is better wherever a tool exists; computer use covers programs with no other interface. Claude itself uses tools first and the screen as a last resort.

How do I turn on Claude computer use on Windows?

In the updated desktop app, under Settings, Desktop app section, General (or Capabilities), turn on Enable computer use. You need a Pro, Max or Ultra plan and a Windows x64 PC.

Does computer use work from my phone?

No: it works through the desktop app open on the PC. To use the PC from your phone you need a remote MCP server added as a connector.

How much does it cost?

In the app it is included in Pro, Max and Ultra, within usage limits. Through the API you pay per token, and every screenshot adds up: a many-step task costs far more than the same task done with MCP tools.

Can they be used together?

Yes, and that is the best choice: MCP tools for files, terminal and apps, computer use as a fallback for what has no other way in.

Further reading

MCP tools for your whole Windows PC, desktop programs included. From any Claude chat, phone included. Seven days free.

Discover MCP DeskManager Ask the assistant