Anthropic Gives Claude Hands: Code and Cowork Now Drive Your PC

Anthropic flipped the switch on computer-use for two of its Claude products Monday. Code and Cowork—until now chat-based helpers—can now grab control of your mouse and keyboard to finish entire workflows. The feature is live as a research preview for paying users on Pro, Max, Team, and Enterprise plans. Expect cursor movement, text entry, and full app launching with the same model that powers Claude 3.5 Sonnet.

Think remote-desktop, but the remote hand belongs to Claude. You grant permission once, then describe a goal—“summarize every PDF in this folder and email the notes to Sarah.” The agent opens Finder or Explorer, clicks through each file, drafts the summary in your mail client, and drops the finished message into your drafts folder. The system uses vision to read the screen and reinforcement learning to decide the next click, key, or menu choice. Sessions stay alive across devices; start on your desktop, check progress from your phone, then review the final doc on your tablet.

Anthropic sandboxes the agent hard. By default it cannot touch crypto or trading apps and refuses credential entry. Prompt-injection filters watch for malicious instructions mid-task, and an allow-list gates which executables can be launched. Even so, the company calls the release a “research preview” and warns that safeguards “aren’t absolute.” Users must opt-in after a security splash screen, and every session can be killed instantly with a single button.

Pressure from OpenClaw, Google’s Project Jarvis, and Microsoft’s Copilot Studio has turned computer-use into the next arms race. Anthropic’s models already ranked high on SWE-bench for coding tasks; giving them literal hands lets the company leap from writing code to shipping entire features. The move also unlocks a new revenue tier: knowledge workers who’ll pay premium prices to offload grunt work like slide polishing, CRM cleanup, or competitor research.

Code users get a turbocharged pair-programmer. Instead of pasting snippets back and forth, Claude can now run the build, open failing tests, patch the source, commit, and push. The same API endpoint that drives Cowork is available to builders, meaning any SaaS can bolt on a “let Claude drive” button. Early testers report 40–60% faster iteration loops, though brittle UIs still trip the agent about 12% of the time.

Cowork targets the docs-and-spreadsheet crowd. Ask it to reconcile two Excel files, pull charts into PowerPoint, and schedule the deck for tomorrow’s meeting. The agent works while you grab coffee, sending a mobile ping when done. Early reviews from Wired call it “the first agent that actually finishes chores instead of hallucinating a plan.” Expect $20–30 monthly add-on pricing once the preview ends.

Screen-scraping malware could hijack the control channel, rogue prompts might pivot from “organize invoices” to “wire money,” and hallucinated clicks could nuke hours of work. Anthropic logs every action locally and hashes sensitive data, but enterprise security teams still want SOC 2 attestations before rolling it out. The company recommends starting with non-production machines and read-only file access until trust builds.

Anthropic will watch telemetry for failure modes, then widen the beta to all paid tiers. A Windows version follows the current Mac release, and deeper integrations with Google Workspace and Microsoft 365 are already in QA. Pricing will shift from research-preview free to usage-based tiers tied to task length and compute time. If adoption spikes, expect rival labs to ship similar features within months, turning computer-use from headline stunt to table stakes overnight.