What is Claude Cowork?
Claude Cowork (often shortened to "Cowork", though the disambiguated form is preferred because of the obvious coworking-space collision) is Anthropic's sandboxed desktop agent runtime — a virtualized environment where Claude can read files, run tools, and execute multi-step work without leaving the user's machine.
It runs locally on macOS using Apple's virtualization framework. The session data — prompts, files touched, tool calls, assistant responses — stays inside the VM on the user's device by default. That's a deliberate privacy decision, and it's the same decision that means Claude Cowork activity does not appear in Anthropic's Audit Logs, Compliance API, or Data Exports. The official admin dashboard surfaces usage metrics (counts, active users) but not content.
For most individual users that's fine. For teams shipping Cowork to clients or rolling it out internally, the absence of a server-side session record creates a real gap whenever someone needs to reconstruct what an agent did on a given day — for QA, compliance, training, or just debugging.
The standard ways to close the gap are:
- OpenTelemetry — emit metrics (cost, latency, tool counts) to your own collector. Aggregate, not content. Explicitly not a compliance-grade audit trail per Anthropic's docs.
- A capture plugin running inside Cowork — installs as a personal or organization plugin, hooks every session event, ships them to a dashboard you control. Argus is one such tool.
In short: Cowork is the agent runtime; the visibility you get into it is whatever you build on top.