Anthropic Gives Claude Computer Use on Mac — and You Can Control It From Your Phone

Anthropic shipped computer use for Claude Code and Claude Cowork on March 23, 2026. The research preview lets Claude point, click, scroll, and type on your Mac — opening apps, navigating browsers, filling spreadsheets, running dev tools — essentially anything you’d do sitting at your desk. Combined with Dispatch, which launched a week earlier, you can text Claude a task from your phone, walk away, and come back to finished work on your laptop.

The desktop agent race just got a lot more crowded.

How Claude Computer Use Actually Works

The system follows a two-tier approach. When you assign Claude a task, it first checks whether a direct connector exists — Google Workspace, Slack, and other supported integrations. If a connector is available, Claude uses it. If not, Claude falls back to GUI control: it takes screenshots of your screen, identifies UI elements, and operates the mouse and keyboard like a human would.

This fallback design is pragmatic. API integrations are faster and more reliable, but they don’t cover every app. By adding screen-level control as a second layer, Anthropic effectively makes Claude compatible with any macOS application without waiting for each developer to build an integration.

For developers specifically, Claude Code can now navigate IDEs, submit pull requests, run test suites, and manage terminal workflows — all autonomously. The idea is that you describe what you want done, and Claude figures out which buttons to click.

Dispatch: The Phone-to-Desktop Bridge

Dispatch, which went live on March 17 as a separate research preview, turns your phone into a remote control for Cowork sessions running on your Mac.

Setup is minimal: open Claude Desktop, start a Cowork session, click Dispatch, scan a QR code with your phone. From there, you can message Claude tasks — “update the Q1 numbers in that Excel sheet,” “check the staging environment for errors” — and it executes them on your desktop.

One important caveat: your Mac has to stay awake. Dispatch is a remote control, not cloud computing. If your laptop sleeps or the app closes, the session stops. This is a meaningful limitation for anyone hoping to fire off tasks and forget about them for hours.

Still, the workflow is compelling. You’re on the train, you remember something needs updating, you text Claude, and it’s done by the time you sit down at your desk.

The Permission-First Safety Model

Giving an AI full control of your computer raises obvious security questions, and Anthropic is leaning hard into a permission-based model to address them.

Claude asks for explicit approval before accessing any new application. Users can stop execution at any time. Read-only permissions are the default — Claude needs your sign-off to edit files, run commands, or take destructive actions. There’s also a command blocklist that blocks risky operations like curl and wget by default, and input sanitization to guard against prompt injection.

Anthropic is also using filesystem isolation (Claude can only access specified directories) and network isolation (only approved server connections) as additional layers.

That said, Anthropic itself acknowledges the limitations. The official recommendation during the research preview: avoid giving Claude access to sensitive data. Security researchers have already flagged that computer-controlling AI agents could theoretically be exploited to download malware or exfiltrate information through prompt injection attacks. The permission system mitigates but doesn’t eliminate these risks — which is exactly why this is labeled a research preview, not a general release.

Where Claude Computer Use Fits in the Desktop Agent Race

Claude isn’t the only AI trying to take over your desktop. The competition is intense and accelerating:

OpenAI Codex has evolved far beyond its 2021 origins. The 2026 version is a full autonomous software engineering agent powered by GPT-5.2-Codex, available as a cloud web agent, an open-source CLI, IDE extensions, and a macOS desktop app launched in February 2026. It scores around 80% on SWE-bench Verified. But Codex is primarily developer-focused — it doesn’t aim to control your entire desktop.

OpenClaw, the open-source darling with 219,000 GitHub stars, runs locally and connects to WhatsApp, Slack, iMessage, and more. It handles practical tasks like clearing inboxes, managing calendars, and running scripts. Free to use (you pay API costs), but setup complexity and security risks come with the territory.

Perplexity Computer, launched February 25, takes a cloud-first approach with 19 AI models and multi-agent orchestration. It’s arguably safer since everything runs in Perplexity’s controlled environment rather than on your local machine. But it costs $200/month and gives you less control.

Manus AI positions itself as the low-barrier option — a cloud sandbox agent that requires no local installation, targeting non-technical business teams.

Claude Computer Use sits in a specific niche: local execution with a permission-first model, integrated into a broader ecosystem (Claude Code for developers, Cowork for knowledge workers, Dispatch for mobile access). The trade-off versus cloud-based alternatives is clear — you get more control and lower latency, but you also accept more responsibility for security.

On benchmarks, Claude Opus 4.6 leads on OSWorld-Verified, which specifically measures GUI navigation and computer use tasks — a direct validation of Anthropic’s approach.

Who Gets Access and What It Costs

Claude Computer Use is available now on macOS as a research preview for:

Claude Pro subscribers at $20/month
Claude Max 5x subscribers at $100/month
Claude Max 20x subscribers at $200/month

Windows and Linux support is not available yet — Anthropic hasn’t announced a timeline.

The Pro plan gets you access to the feature, but Max subscribers get significantly higher usage limits (5x or 20x more capacity), which matters if you plan to use computer control extensively. Heavy automation tasks burn through usage faster than chat conversations.

What’s Still Missing

This is a research preview, and it shows. A few notable gaps:

Platform lock-in. Mac-only for now. A large chunk of developers and enterprise users on Windows or Linux are locked out.

Sleep dependency. Your Mac must stay awake for Dispatch to work. No background processing, no cloud fallback.

Speed. GUI-based control is inherently slower than API calls. Claude has to take screenshots, process them, identify elements, and then act. For tasks where a direct integration exists, the connector path is much faster.

Trust boundary. Even with permissions, giving an AI the ability to click around your desktop requires a level of trust that many users and IT departments aren’t ready for. The “research preview” label gives Anthropic cover, but it also signals that even they consider this experimental.

FAQ

Is Claude Computer Use free?
No. It requires a Claude Pro ($20/month) or Claude Max ($100-$200/month) subscription. There is no free tier access.

Does Claude Computer Use work on Windows or Linux?
Not yet. The research preview is macOS-only. Anthropic hasn’t announced when other platforms will be supported.

How does Claude Computer Use compare to OpenClaw?
OpenClaw is open-source, free (minus API costs), and supports more platforms, but requires technical setup and runs with fewer safety guardrails. Claude Computer Use is a managed experience with built-in permissions and a polished mobile companion (Dispatch), but it’s Mac-only and requires a paid subscription.

Can Claude access my passwords or sensitive files?
Claude operates under a permission-first model and asks before accessing new applications. However, Anthropic recommends avoiding sensitive data during the research preview. The filesystem isolation limits Claude to specified directories.

What is Claude Dispatch?
Dispatch is a companion feature that lets you assign tasks to Claude from your phone. It launched on March 17, 2026, and works by connecting your mobile device to an active Cowork session on your Mac via QR code. Your computer must remain awake for it to function.

Top AI Product

Leave a comment Cancel reply