Claude Code Desktop gains “computer use” to control your Mac

Anthropic is bringing direct “computer use” control to Claude Code Desktop, letting the AI operate apps via mouse-and-keyboard actions. Early reactions praise the workflow shift but raise sharp questions about permissions, guardrails, and platform support.

Claude Code Desktop gains “computer use” to control your Mac

TL;DR

  • Computer use control: Claude Code can act via mouse, keyboard, and app-level UI
  • Availability: Limited to Claude Code Desktop, not the broader Claude experience
  • Workflow shift: Prompting becomes delegation; UI/screen context joins the feedback loop
  • Security concerns: Questions around permissions, access controls, and app approval granularity
  • DIY reference: CDP + Playwright MCP noted as workable, but prone to destructive actions without tight constraints
  • Pace and platforms: Calls for Windows/Linux support; claim acquired a “computer use company” and shipped four weeks later

Anthropic’s Claude Code is picking up a new capability—direct “computer use” control—and the first detail that seems to matter in practice is where it’s showing up: Claude Code Desktop, not the broader Claude experience.

Boris Cherny says a small Anthropic Labs crew was responsible for several of the building blocks that now define Claude’s developer-facing ecosystem—work that culminates, at least in this thread, with full computer use available from Cowork and Dispatch.

Cherny also points back to early “Desktop computer use prototypes” from the “Sonnet 3.6 days,” describing them as “clunky and slow,” but still legible as a bet: ship a surface early, and let model capability grow into it.

“Full computer use” arrives in Cowork and Dispatch

Other replies focus on day-to-day friction points that tend to surface once a tool is used continuously: requests for multi-window support in Cowork, app stability issues (including a report that “the connection with Chrome keeps breaking”), and repeated asks for Linux availability for Claude’s desktop-oriented experiences.

What “computer control” changes for AI-assisted coding workflows

That distinction came through in community reactions. Kai (@ZerosByKai) framed computer control as more than a demo-friendly feature, arguing it effectively turns prompting into delegation—“a junior dev who sits at your actual desk.”

Guardrails, permissions, and the uncomfortable questions

Alongside the excitement, the replies quickly converged on risk. Multiple posters asked about permissions and access controls, including what happens when Claude interacts with apps that haven’t been explicitly approved, and how granular controls are “in practice.”

Users noted that DIY approaches using CDP + Playwright MCP can work, but also reveal how easily an agent can take destructive actions without tight constraints—like closing browser tabs unless explicitly told not to. Artem Rybachuk pushed on the broader trust issue: after years of locking computers down, it’s jarring to “willingly hand control over to AI,” especially with personal data in the mix.

Original source

Continue the conversation on Slack

Did this article spark your interest? Join our community of experts and enthusiasts to dive deeper, ask questions, and share your ideas.

Join our community