Claude PC Control | Comparing Cowork, Code, and Computer Use
For developers who want to control a PC with Claude, this article provides a thorough comparison of the three paths — Cowork, Claude Code Computer Use, and Computer Use Tool — based on the latest May 2026 specifications. It covers supported operating systems, required plans, differences in permission flows, and how to choose based on your use case, all grounded in official Anthropic information, presented in an order that helps you select a path as quickly as possible.
The easiest path to get started is Cowork. It supports both macOS and Windows, and can be launched immediately by installing Claude Desktop on a Pro or Max plan — making it the first choice for PC control as of May 2026 if you're unsure where to begin.
The developer CLI path is Claude Code Computer Use, which is macOS-only and activated via /mcp to enable computer-use in an interactive session running Claude Code v2.1.85 or later. Its primary use cases are end-to-end UI testing and GUI verification — it is not intended for tasks that can be handled through text alone.
For OS-independent automation, choose the Computer Use Tool (API) and call computer_20250124. However, providing the screen and dispatching input requires your own sandbox implementation, and starting from Anthropic's official reference Dockerfile is the safe approach.
目次 (8)
- What Is Claude PC Control — A Research Preview That Sees and Controls the Screen
- The 3 PC Control Paths — Cowork, Claude Code, and Computer Use Tool
- Controlling Your PC with Cowork — The Easiest First Path
- Controlling Your PC with Claude Code Computer Use — The Developer CLI Path
- Controlling Your PC with Computer Use Tool (API) — The Self-Hosted Sandbox Path
- Dispatch — Having Claude Work While You're Away from Your PC
- Security and Permission Flow — Explicit Per-App Permission Is Required
- How to Choose by Use Case — Which Path Is Right for You?
What Is Claude PC Control — A Research Preview That Sees and Controls the Screen
Claude PC control is a research preview feature that allows Claude to "see" the screen through screenshots, operate the mouse and keyboard, and autonomously carry out tasks across multiple applications. Anthropic announced it to the public on March 23, 2026, after which it was rolled out across three paths: Cowork, Claude Code, and the API (source: Anthropic says Claude can now use your computer).
Claude captures pixel information from the screen as screenshots, visually interprets button positions, text fields, and menu items, and returns actions accordingly. Because it does not rely on app-specific APIs, virtually anything a human can operate through a GUI — native apps, web browsers, terminals, spreadsheets — is in scope in principle. At the same time, Anthropic itself notes that this is "still an early-stage capability compared to text or code tasks," which means designs that account for the risk of misclicks and mistypes are required (source: Put Claude to work on your computer).
The 3 PC Control Paths — Cowork, Claude Code, and Computer Use Tool
To help you understand the differences between the three paths at a glance, the table below summarizes supported OS, required plan, prerequisites, permission flow, Dispatch support, and primary use cases (each cell reflects the facts explained in the body of this article).
| Path | Supported OS | Required Plan | Version / Prerequisites | Permission Flow | Dispatch Support | Primary Use Cases |
|---|---|---|---|---|---|---|
| Cowork | macOS / Windows | Pro / Max | Claude Desktop app | Explicit per-app permission + OS permissions (Accessibility & Screen Recording) | Supported (macOS only) | Lowest-barrier first path; supervised pair work |
| Claude Code Computer Use | macOS only | Pro / Max | Claude Code v2.1.85+; interactive session; computer-use enabled via /mcp |
Explicit per-app permission + OS permissions (Accessibility & Screen Recording) | — (not documented) | E2E UI testing; native app build verification; visual bug reproduction |
| Computer Use Tool (API) | OS-independent | — (via API) | Pass computer_20250124 to Messages API; own sandbox (reference Dockerfile) required |
No GUI permission dialog; isolation design is developer's responsibility | — (not documented) | RPA-style parallel deployment in production sandboxes; Linux environments |
All three paths differ only in "where Claude is called from" — the core visual-to-action conversion is common across them. Cowork delegates screen permissions to Claude from Claude Desktop's dedicated UI; Claude Code does so from a developer CLI session; Computer Use Tool does so from an Anthropic API client.
The supported OS and usage conditions differ significantly by path. Cowork supports both macOS and Windows and is available to Pro and Max plan subscribers simply by installing Claude Desktop. Claude Code Computer Use has four constraints — macOS only, Pro or Max plan, Claude Code v2.1.85 or later, and interactive session only — and only activates in projects where computer-use has been enabled via /mcp (source: Let Claude use your computer from the CLI). Computer Use Tool (API) is OS-independent, but screen provision and input dispatch must be implemented by the developer; Anthropic provides a reference sandbox Dockerfile as a starting point (source: Computer use tool — Claude API Docs).
Controlling Your PC with Cowork — The Easiest First Path
Cowork is a feature that grants Claude desktop control permissions from the Claude Desktop app, and it is the lowest-barrier option among the three paths. Log in to Claude.com, install Claude Desktop on macOS or Windows, start a Cowork session, and Claude will operate the GUI while explicitly requesting permission for each individual application (source: Let Claude use your computer in Cowork).
Supported plans are Pro and Max only; it is not available on the Free plan or certain Team / Enterprise configurations. Once you grant OS-level Accessibility and Screen Recording permissions the first time, Claude then operates in two layers: it checks in app-by-app, asking "May I access Chrome?" or "May I access Numbers?" as it goes. Because this is a research preview, the recommended safe usage pattern is supervised "pair work" — with the user watching alongside — rather than extended unattended automation (source: Computer use tool — Claude API Docs (safety guidance); supplemental: Claude Computer Use Security Risks 2026 Guide).
Controlling Your PC with Claude Code Computer Use — The Developer CLI Path
Claude Code Computer Use is the path that delegates macOS screen control to Claude from a Claude Code CLI session. It provides a built-in MCP server called computer-use, which is disabled by default and only activates in projects where it has been enabled via the /mcp menu. The four usage requirements are: macOS only, Pro or Max plan, Claude Code v2.1.85 or later, and an interactive session (-p non-interactive mode is not supported) (source: Let Claude use your computer from the CLI).
Anthropic positions Computer Use as "the broadest and slowest tool," documenting it within Claude Code as the tool to reach for last, after MCP → Bash → Claude in Chrome → Computer Use in that order. Its primary domain is tasks that "cannot be judged without actually seeing the screen" — such as end-to-end UI testing, native app build verification, and visual bug reproduction — and the recommended pattern is not to use it for tasks that can be handled through text. Once you grant the two macOS permissions — Accessibility and Screen Recording — setup itself takes about five minutes.
Controlling Your PC with Computer Use Tool (API) — The Self-Hosted Sandbox Path
Computer Use Tool is the primary interface provided on the Claude API side. Passing the tool identifier computer_20250124 to the Messages API activates it. Claude returns actions such as mouse_move, left_click, type, key, and screenshot; the developer's runtime then executes those on its own sandbox and returns the resulting screenshot as the next turn, forming a loop (source: Computer use tool — Claude API Docs).
Anthropic's official reference implementation runs a Linux desktop inside a Docker container and handles screen provision and input dispatch via VNC. Because it is sandbox-based, the strength of the API path is its ability to move into production use cases — such as building a Windows VM running internal systems in an RPA-like manner, or running multiple Claude agents in parallel on ECS Fargate. However, since there is no app-level permission dialog as there is with Cowork or Claude Code, all isolation design for the target environment is the developer's responsibility.
Dispatch — Having Claude Work While You're Away from Your PC
Dispatch is an "autonomous execution while away" feature used in conjunction with Cowork. It allows Claude to continue a task even after the user has stepped away from the PC, with results waiting as a report upon return. Anthropic announced Dispatch and Computer Use simultaneously at the March 2026 launch, framing it as "an AI colleague that keeps working while you're out to lunch" (source: Claude March 2026: Computer Use, Dispatch and Anthropic news).
As of March 2026, Dispatch is macOS-only; the path to use Dispatch alongside the Windows version of Cowork has not yet been added to the rollout roadmap. When using Dispatch, the prompt design that tells Claude "which apps it may touch" and "how far to proceed before stopping" is critical to safety. Because Anthropic itself has explicitly stated that it is continuing to improve safety mechanisms, it is a practical judgment to avoid assigning irreversible operations — such as payments, production deployments, or deletions — for the time being.
Security and Permission Flow — Explicit Per-App Permission Is Required
The safety mechanism common to all three paths is the explicit per-app permission flow. With Cowork and Claude Code Computer Use, every time Claude accesses a new application, a popup asks "Do you want to allow operations in this app?" Claude cannot touch that application's window unless the user approves (source: Let Claude use your computer in Cowork).
In addition, OS-layer permissions for Screen Recording and Accessibility are prerequisites. On macOS, you must grant both permissions to Claude Desktop / Claude Code via System Settings > Privacy & Security; forgetting to do so results in a state where "Claude appears to be operating but nothing happens." Anthropic also explicitly identifies the risk of malicious instructions being injected via prompt injection, and when operating a web browser, using login isolation (profile separation or guest mode) in combination is the recommended pattern (source: Computer use tool — Claude API Docs (prompt injection countermeasures); supplemental: Claude Computer Use Security Risks 2026 Guide).
How to Choose by Use Case — Which Path Is Right for You?
Path selection comes down to three axes: "OS," "developer vs. non-developer," and "degree of automation autonomy." If you are a non-developer who wants to get started easily on either macOS or Windows, Cowork is the first candidate. If you are a developer on macOS who wants to incorporate UI operations within a Claude Code session, Claude Code Computer Use is the option; if you want to deploy in parallel RPA-style in a production sandbox, Computer Use Tool (API) is the choice.
For users who want Claude to control a PC in a Linux environment, the only path officially supported at this time is implementing Computer Use Tool (API) with a self-hosted sandbox. Neither Cowork nor Claude Code Computer Use has Linux support on its roadmap, making the practical solution for Linux users to build a Docker sandbox via the API path (source: Computer use tool — Claude API Docs). Since Dispatch-enabled unattended automation is only supported in macOS Cowork, choosing macOS is the fastest route as of May 2026 if you want to set up long-running, hands-off workflows.