Anthropic has released Claude Cowork, a research preview that repositions Claude as an autonomous desktop agent rather than a conversational assistant. Available to paid Claude subscribers on Windows and macOS through Claude Desktop, it runs inside a sandboxed Linux virtual machine on the user's own machine — keeping code execution local while routing prompts and file contents to Anthropic's cloud for inference.
The architecture is deliberate about that split. The sandbox provides a security perimeter, and users grant access at the folder level rather than giving the agent free run of their file system. But anything the agent needs to reason about — including document contents — leaves the machine. Enterprise buyers in regulated industries will want to factor that in before deployment.
External integrations are notably absent out of the box. Connections to Gmail, Slack, Google Drive, and Notion must be sourced, installed, and authenticated independently by users via the Model Context Protocol (MCP), the open standard Anthropic has been pushing across its agent products. It's a flexible model, but it shifts setup burden onto users and produces a more fragmented experience than platforms that bundle integrations natively.
Scheduled automation is supported — daily briefings, weekly report compilations — but only while the host machine and Claude Desktop remain active. That limits Cowork's appeal for teams expecting continuous background processing, and marks a clear distinction from server-side agent platforms that can run unattended in the cloud.
Anthropoc is positioning Cowork away from Claude Code, its developer-facing offering, and toward non-technical knowledge workers: sales teams, marketers, data analysts, project managers, finance staff, researchers. Persistent global and folder-level instructions let users encode role context, preferred tone, and hard constraints that carry across sessions — a practical alternative to re-engineering each prompt from scratch.
A third-party guide published by Overton Collective on March 10 provides the most detailed public breakdown of the product to date, including downloadable template packs for common professional roles. Its existence points to a nascent ecosystem forming around Anthropic's agentic products, even while Cowork itself remains in research preview.
The competitive picture is familiar. OpenAI has Operator and a growing set of desktop agent capabilities; Google is expanding agentic functionality across Workspace. The user's own machine is fast becoming the primary battleground for AI platform lock-in, with each major lab working to establish their model as the default background worker. Whether Cowork's sandboxed VM and MCP plugin model can satisfy enterprise expectations around auditability and continuous availability will determine how far it travels once it exits preview.