Two days ago, Anthropic quietly made Claude Cowork available to all paid subscribers, and it represents the biggest functional shift in Claude since the introduction of Projects.
Cowork moves Claude from conversation partner to autonomous worker. Instead of answering questions in a chat window, it operates inside a lightweight Linux virtual machine running directly on your computer. It can read and write files, execute Python, JavaScript, and shell commands, and chain together multi-step tasks without you babysitting each one.
How It Actually Works
The architecture splits neatly in two. The LLM inference (the "thinking" part) still happens on Anthropic's servers. But execution happens locally in a sandboxed VM on your machine. Claude can only access folders you explicitly grant permission to, so it is not rummaging through your entire hard drive.
The practical upside: you can point Cowork at a messy folder of PDFs and say "organize these by client and summarize each one," and it will actually do it. File management, research tasks, document drafting, data analysis, and coding are where it performs best.
For connecting to external services like Gmail, Slack, or Google Drive, Cowork uses Model Context Protocol (MCP) integrations. These require separate setup and authentication, but once configured, Claude can send emails, post messages, or pull files from cloud storage as part of a larger task.
The Catches
Your computer needs to stay awake with Claude Desktop open for tasks to run. There are no audit logs yet, which rules it out for anyone working under HIPAA, SOC 2, or similar compliance requirements. And while Anthropic advertises scheduled recurring automation, the always-on requirement makes that less practical than it sounds.
Who Should Try It
Cowork requires a Claude Pro subscription ($20/month) at minimum, and works on both macOS and Windows through the Claude Desktop app. No technical background needed to get started.
For solo operators and small teams drowning in repetitive file and document work, this is worth testing immediately. The sandboxed approach feels meaningfully safer than giving an AI agent full system access, which is what most competing tools require. The MCP integration layer is still clunky to set up, but once running, the ability to chain together "pull data from Drive, analyze it, draft a summary, email it to the team" into a single autonomous task is genuinely useful.