Anthropic has rolled out autonomous computer-control to its Claude Code and Cowork tools, allowing the AI to directly operate a user's Mac — opening files, running browsers and applications, and executing developer tools without manual configuration.
The feature is available now as a research preview for subscribers on Claude's Pro and Max tiers, according to Anthropic's announcement. Computer control is currently limited to macOS, with broader platform support unconfirmed.
From Developer API to Consumer Product
The new capability builds directly on computer-use features Anthropic first introduced through Claude 3.5 Sonnet in 2024 — a milestone that allowed Claude to interpret screen contents and simulate mouse and keyboard inputs. That earlier rollout was accessible primarily through developer APIs and required technical configuration. Bringing the same functionality into Code and Cowork, with minimal setup for paying subscribers, marks a shift toward mainstream deployment of what Anthropic itself classifies as a research-stage capability.
The 'no setup required' framing signals a deliberate move to bring autonomous computer control out of developer sandboxes and into everyday subscriber workflows.
Claude Code targets software development workflows; Cowork is designed for broader collaborative task completion. Integrating autonomous computer control into both products suggests Anthropic is pursuing AI assistance across a wide range of desktop tasks — not solely coding.
How the Permission Model Works
Before taking any independent action, Claude requests explicit user permission, including for tasks initiated while users are away from their computers. Anthropic frames this as assistive rather than fully unsupervised — though the degree of ongoing human oversight will vary depending on the task and user preference.
The permission requirement is one safeguard, but Anthropic has not disclosed what logging or audit capabilities are available to users who want to review actions Claude took during an autonomous session. The company has also not published accuracy benchmarks for the updated implementation in Code or Cowork.
A Different Go-to-Market Approach Than Rivals
Autonomous computer control has become a competitive priority among major AI developers. OpenAI launched its Operator product earlier in 2025, enabling browser and application automation through ChatGPT. Google has pursued comparable functionality through its Project Mariner research effort.
Anthropic's decision to embed the feature within existing subscriber tools — rather than launch it as a standalone product — reflects a distinct strategy. Users who already pay for Pro or Max access can try the capability without signing up for anything new.
Security concerns have accompanied every major computer-use announcement in the sector. When Anthropic first demonstrated computer use in 2024, the company acknowledged the model could make mistakes, misinterpret screen contents, or take unintended actions. Some security researchers have called for more granular approval mechanisms when AI systems interact directly with file systems and applications — the current permission model does not go that far.
The research preview designation indicates Anthropic does not consider the feature at full production quality. The company has used similar framing before wider releases, typically using preview periods to collect feedback and address failure modes at scale. No timeline has been announced for expanding support to Windows or other platforms, nor for lifting the research preview status.
What This Means
For professionals who spend significant time on repetitive desktop tasks, autonomous computer control in Claude Code and Cowork represents a meaningful capability upgrade — but users should treat it as an early-stage tool and review what actions Claude takes on their behalf until Anthropic publishes clearer reliability and audit standards.