On March 23, 2026, Anthropic shipped a feature that redefines what an AI assistant can be: Computer Use. For the first time, Claude can open apps, move the cursor, type text, navigate browsers, and complete tasks on your Mac — all while you're away from the desk.
This is not a gimmick. It is the direct continuation of a product philosophy Anthropic has been building toward since Remote Control (February 2026) and Channels (March 2026): Claude as an autonomous agent that works when and where you need it, not only when you're sitting in front of a terminal.
How It Works
The architecture follows a smart priority system. When you assign Claude a task — through Dispatch on your phone, or directly in Claude Code or Cowork — it first checks whether it has a direct connector for the apps involved. If Gmail, Slack, Google Calendar, or another integrated service is part of the task, Claude calls the relevant APIs directly. No screen involved, no latency from visual navigation.
Only if no connector exists does Claude fall back to screen control: it takes a screenshot, analyses what it sees, then clicks, scrolls, and types to navigate the application like a human would.
This distinction is deliberate and important. Screen-based navigation is slower and more error-prone than direct API calls. By positioning it as the fallback rather than the default, Anthropic has built a system that degrades gracefully rather than relying on fragile pixel-level automation.
The Dispatch Connection
Computer Use launched alongside Dispatch, a feature released the week before that lets you converse with Claude across your phone and desktop. The pairing is the point: you assign Claude a task from your iPhone while commuting, and return to finished work on your desktop.
An Anthropic demo shows exactly this: a user asks Claude to export a pitch deck as a PDF and attach it to a meeting invite — via a phone message, while away from the Mac. The Mac runs in the background, Claude opens the relevant apps, does the work, and the result is waiting when you return.
Setup and Requirements
Computer Use is a research preview, available to Pro and Max subscribers on macOS. It is not available on Team or Enterprise plans, and Windows support is confirmed but not yet scheduled.
To enable it:
- Open Settings → Desktop app → General and toggle Computer Use on
- Grant two macOS permissions: Accessibility and Screen Recording
- The first time Claude needs to access a specific app, it asks for per-app approval
Approvals are scoped to the current session (or 30 minutes in Dispatch-spawned sessions), which limits the blast radius of any mistake.
What Works — and What Doesn't Yet
First-look reports from March 23 onward paint a consistent picture: Computer Use works reliably for clear, bounded tasks — compiling data from a web page into a spreadsheet, opening a file and attaching it to an email, navigating a GUI that has no CLI alternative. Anthropic's own statement is unusually candid:
"Computer use is still early compared to Claude's ability to code or interact with text."
In practice, MacStories' first-look testing put real-world reliability for complex tasks at roughly 50/50. Multi-step chains with branching decisions are unreliable at this stage. The screenshot-analyse-act cycle also adds latency — each visual step takes a few seconds, which adds up for longer workflows.
Current hard limitations: macOS only (Windows confirmed, no timeline); Mac must be awake with Claude Desktop running; one continuous thread, no parallel task management; not available on Team or Enterprise plans; sensitive information should be avoided during the preview.
The Security Question
Giving an AI agent access to your desktop is not a decision to take lightly. Anthropic is explicit about the risks: screen-based automation opens the door to prompt injection attacks via content on the screen — a malicious web page could theoretically instruct Claude to take actions the user didn't intend.
The mitigations are sensible. Per-app, per-session permission gates mean Claude cannot silently expand its access. The connection runs over outbound HTTPS only — no inbound ports are opened. And the explicit guidance to avoid sensitive information during the preview sets the right expectations.
For developers already comfortable with Claude Code and Remote Control, the risk profile is familiar: you are expanding Claude's autonomy incrementally, with explicit approval gates at each step.
Why This Matters
The combination of Dispatch, Remote Control, Channels, and now Computer Use means Claude has, in the space of five weeks, transformed from a terminal tool into something closer to a background worker that lives in the channels and apps you already use.
For developers, the most immediate value is desktop automation for tools with no CLI or API — legacy internal apps, design tools, browser-based dashboards that predate proper APIs. For anyone drowning in admin work, the pitch is simpler: you assign the task, Claude does the clicking.
The research preview label is honest. This is not finished. But the architecture — connectors first, screen control as fallback, per-app approvals, outbound-only networking — is already more thoughtfully designed than most early computer-use implementations in the market.