Anthropic just opened a research preview of Claude Duet, a standalone app that turns Claude into a proactive agent instead of a plain conversational partner. A quick rundown for curious devs.
What actually changes
Until now, Claude waited for you to write to it. Duet runs continuously on its own virtual machine in the cloud: it can keep working on a task after you close the app, then come back with a result or a request for approval before acting (no silent execution of sensitive actions).
Under the hood, Duet reuses the existing infrastructure of a Claude account:
- Memory and shared conversation history
- Skills already configured
- Connectors: access to over 6,000 integrations from the Claude Marketplace
- Underlying model: Claude Opus 5.5
Security guardrails
Anthropic is emphatic about this: a dedicated classifier and a specific system prompt screen every message to limit sensitive-information leaks. Duet also relies on the same trustworthy-agent framework used by Claude Code and Claude in Chrome, backed by adversarial infrastructure testing. No access to passwords or logins on connected accounts, and no financial actions (trading, transfers).
Acknowledged limitations
Duet isn't built for personal errands (booking a ticket, for instance), and Anthropic openly admits it can still "misread a thread" or miscalibrate urgency, fair enough for a v1.
Availability
Waitlist open for Pro and Max subscribers, with iOS/Android apps and web access, usage counted against existing plan quotas.
Worth watching whether this "agent that comes to you" approach holds up against Muse (Meta) and Dots (OpenAI).
Top comments (0)