Anthropic Brings Claude Desktop Control to Mac

Anthropic has expanded Claude’s capabilities to include direct desktop control on macOS, letting the AI click, type, and navigate applications on a user’s computer when standard integrations with tools like Slack and calendar apps are unavailable.

The feature is available now as a research preview inside Claude Cowork and Claude Code for Claude Pro subscribers ($20/month) and Claude Max subscribers ($100/month). It is currently limited to macOS.

How Desktop Control Works

Claude does not default to desktop control. The system first attempts to complete tasks through existing app integrations — connected calendars, Slack, or other supported tools. Only when no integration is available does Claude take over the mouse and keyboard to complete the task directly.

Users must explicitly approve each application Claude is allowed to control, and they can stop Claude at any point during a session. The design reflects Anthropic’s stated priority of keeping humans in the loop during the research preview phase.

Alongside desktop control, Anthropic released Dispatch — a companion feature that lets users assign tasks to Claude from their iPhone and return to completed work on their desktop. The mobile app connects to an active Cowork session, with all processing happening locally on the Mac.

Vercept Acquisition Underpins the Technology

The capability builds on technology from Vercept AI, a Seattle-based startup Anthropic acquired in February 2026. Vercept was founded in November 2024 by former Allen Institute for AI researchers Kiana Ehsani, Luca Weihs, Ross Girshick, and Matt Deitke, and had raised $16 million in a seed round led by Fifty Years before the deal closed. Angel investors included former Google CEO Eric Schmidt, Google DeepMind chief scientist Jeff Dean, Cruise founder Kyle Vogt, and Dropbox co-founder Arash Ferdowsi.

Vercept’s product, Vy, operated as a cloud-based computer control agent for Mac. Anthropic gave Vercept users 30 days’ notice before shutting down Vy on March 25, and absorbed the team. Co-founder Ehsani said her team “shipped its first product less than four weeks after joining Anthropic,” pointing to Dispatch as the result of that sprint.

Claude’s computer use performance has improved substantially since Anthropic first introduced the capability in late 2024. On OSWorld Verified — a standard evaluation for AI models that operate computers — Claude Sonnet models scored under 15% at the initial release. With Claude Sonnet 4.6, that figure reached 72.5%, approaching the human baseline of roughly 70–75% on the same benchmark. Opus 4.6 scores 72.7% on the same test.

Security and Reliability Remain Open Questions

Desktop control creates risks that browser-based or API integrations do not. When Claude operates a live desktop, it gains access to any visible data on screen, open files, and applications the user has approved. Anthropic’s prior computer use research identified prompt injection — where malicious instructions embedded in on-screen content could override user intent — as a primary attack surface, and Sonnet 4.6 shows improved resistance to this compared to earlier versions.

Reliability is an additional concern. Claude’s earlier computer use architecture processed sequential screenshots rather than a continuous video feed, causing it to miss short-lived actions or notifications. Whether the Vercept team’s approach resolves this at scale is not yet documented publicly.

OpenAI’s own macOS browser-based agent and similar products have faced adoption friction, partly because desktop automation breaks easily when software interfaces change. Anthropic’s choice to limit desktop control to a fallback position — activated only when standard integrations fail — reduces the surface area for errors but also narrows the feature’s scope compared to a fully autonomous desktop agent.

The larger question is whether AI agents for business automation can move reliably from controlled demos to live production environments. Claude’s computer use benchmark score has improved fivefold in roughly 16 months, which is a meaningful trajectory. But benchmark gains on curated tasks do not always predict performance in the messier reality of live macOS environments — unexpected dialogs, application updates, and state changes that no evaluation fully captures. How Anthropic handles security disclosures once researchers begin probing the feature more systematically will be the first real test of whether desktop control is ready to leave its research preview status behind.