Anthropic Adds Desktop Control to Claude Apps

Anthropic added desktop computer control to Claude on March 23, 2026, letting Claude operate your Mac directly inside Claude Cowork and Claude Code. The new Dispatch and computer use release matters because it turns Claude from a connector-driven assistant into a desktop agent that can fall back to clicking through real apps when no structured integration exists.

This is a product release, not a new model launch. Anthropic shipped it as a research preview for Claude Pro and Claude Max subscribers, with macOS support at launch. Dispatch ties the workflow together by letting you assign a task from your phone, then have Claude complete it on your desktop in the same persistent thread.

Product scope

Claude now works across three layers of execution.

First, it uses connectors when available, including Slack, Gmail, Google Drive, and Google Calendar. Second, it can work through the browser. Third, when neither path is sufficient, it can use computer use to point, click, type, open files, navigate apps, and interact with developer tools on your actual machine.

That ordering matters. Anthropic is treating desktop control as the fallback path, not the preferred one. If you build agents yourself, this is the same design pressure behind function calling and structured tools: use precise interfaces first, then drop to UI automation only when necessary.

Dispatch changes the execution model

Dispatch is the orchestration layer that makes the release more than a local desktop feature. You can assign a task from your phone, Claude chooses the right environment, and the work continues on desktop. Development tasks route into Claude Code. Knowledge work routes into Cowork. Results and follow-ups return to one continuous Dispatch conversation, with push notifications when a task finishes or needs approval.

For developers building agents, this is a concrete example of cross-device state management. The task is no longer bound to one chat surface. It persists across mobile initiation, desktop execution, and app-specific approvals, which is closely related to the production concerns in agent memory and stateful agents.

Security boundaries are tighter, and more important

Computer use in Cowork runs outside the virtual machine Cowork normally uses. Claude is interacting with your live desktop, not just an isolated sandbox. Anthropic added per-app permissions, an app blocklist, and action review intended to detect signs of prompt injection.

Some app categories are blocked by default, including investment and trading platforms and cryptocurrency apps. Claude also asks for permission before accessing each application, and you can stop it at any time.

This is the main technical tradeoff. Desktop agents are more capable because they can work anywhere a user can click. They are also harder to secure because the interface is visual, indirect, and exposed to hostile content on screen. If your team is evaluating agent systems, prompt-injection resilience is not a side issue. It is part of the core runtime, as recent work on agent evaluation and monitoring keeps reinforcing.

Screenshots become part of the context

Claude uses screenshots of your computer to understand screen state and navigation. Anthropic is explicit about the implication: anything visible on screen may be exposed to the model, including personal data, sensitive documents, or other people’s information.

Cowork also retains memory across sessions, while excluding categories such as passwords, financial details, and health information from memory. Users can view, edit, and delete remembered information.

If you work on long-running agent workflows, this is a practical reminder that context engineering is no longer just prompt assembly. It includes screen state, local files, app permissions, and memory retention rules. The design problem looks much closer to operating-system mediation than chat UX, which is why context engineering matters more as agents move onto the desktop.

Availability and launch constraints

The launch details are narrow enough to matter in procurement and rollout planning.

Capability	Launch status
Computer use	Research preview
Eligible plans	Pro, Max
Launch OS support	macOS
Team / Enterprise access	Not available for computer use at this time
Desktop requirement	Claude Desktop app installed and running
Mobile requirement	Claude mobile app updated
Threading model	One continuous Dispatch thread

Anthropic also notes operational constraints: your desktop must stay awake and active, the Claude Desktop app must remain open, complex tasks may need a second try, and screen interaction is slower and more error-prone than connectors.

Support documentation around Dispatch already references the latest desktop app on macOS or Windows x64, while computer use itself remains positioned as macOS-first with Windows support still in progress. For teams planning pilots, the safe assumption is simple: Dispatch spans mobile and desktop, but live screen control launched on Mac.

Comparison with Claude’s existing agent surfaces

Anthropic already exposed a developer-facing computer use tool in Claude Docs, with actions such as screenshot, left_click, type, key, scroll, and wait. The difference in this release is product integration. Claude is now executing those capabilities through the desktop app on your machine, inside Cowork and Claude Code, instead of leaving tool execution entirely to an external app developer.

This also separates the feature from Claude Code on the web, which runs asynchronously on Anthropic-managed virtual machines tied to GitHub repositories. Desktop computer use targets local files, installed apps, and the live user environment. If your workflow depends on local IDE state, native apps, or machine-specific context, this is a different category of agent than cloud-only coding systems discussed in coding assistant comparisons.

If you plan to use Claude’s new desktop control, treat connectors as the default, keep sensitive apps out of reach, and design approval checkpoints around every action that crosses from structured tools into live screen interaction.