Claude Duet
Anthropic’s proactive agent: morning briefs and background tasks on its own cloud computer. Pro/Max/Ultra, US research preview.
Group: Work and business agents
Claude Duet is Anthropic’s proactive agent: you hand off work by text or voice message, it keeps working on its own cloud computer while you are away, and it comes back with finished docs, sheets, decks or dashboards - or asks when it needs your approval. It shares your Claude account’s memory, skills and 6,000+ connectors.
The decision in plain English
No extra charge beyond a Claude Pro ($20/month), Max (from $100/month) or Ultra subscription, but it is a US-only research preview reached through a waitlist, and its background work counts against the same plan usage limits. Approval-first design with hard limits on money movement. Output quality untested.
Best fit
Professionals and small-business owners already paying for Claude who want an agent that briefs them, chases work in the background and drafts for approval.
Poor fit
Anyone outside the US, on a free or Team/Enterprise plan, or needing errands, bookings or live browser transparency - Anthropic says Duet is not built for those.
Documented capabilities
Vendor statements, not verified task outcomes.
- Proactive morning brief and attention-flagging across connected inbox and calendar Source · checked 2026-10-26
- Background task execution on its own cloud computer, returning docs, sheets, decks and dashboards Source · checked 2026-10-26
- Small-business operations: finding unpaid invoices and drafting reminders, payroll preparation for review, contract preparation Source · checked 2026-10-26
- 6,000+ connectors and plugins via the Claude Marketplace Source · checked 2026-10-26
- Purchases with one-time merchant-locked wallet cards, approved by the user in the wallet app Source · checked 2026-10-26
Permissions and human control
- Approval-first for sensitive actions: before sending a message, paying or deleting, it shows the action, destination and content including every recipient, and waits (vendor safety page, read 26 October 2026).
- Some actions can never skip approval (paying or sending money, deleting an order, creating an account); securities trading, inter-account transfers and accessing others’ accounts are off-limits entirely.
- Purchases use one-time merchant-and-amount-locked cards from a wallet such as Link from Stripe, approved in the wallet app; the real card number is never shared with Duet or the merchant.
- Passwords, cookies and sign-in tokens never reach the model; they pass through separate machines the model cannot access.
Documented controls do not prove that an agent always follows them.
Limits and open questions
- US-only research preview behind a waitlist; availability and behaviour may change.
- Background checks and larger tasks consume the same Claude plan usage limits as conversations.
- Cannot trade securities, move money between accounts or access other people’s accounts, even with approval.
- Anthropic acknowledges its outbound-message safety classifier will occasionally miss something or over-flag during the preview.
- Output quality, task reliability and approval cadence were not tested.
Privacy and security
Vendor safety page (read 26 October 2026): approvals and safety checks run outside the model on separate machines; an outbound-message classifier screens drafts, with “never share” treatment for credentials, card and ID numbers, and consent-based sharing for health or finance details. Conversation history and memory persist until the user deletes them; deleting the account wipes everything Duet stored. Anthropic cites external penetration testing and red-teaming of credential features. No statement on model training was found on the Duet safety page; the general Claude privacy terms apply.
Common questions about Claude Duet
What is Claude Duet for?
Professionals and small-business owners already paying for Claude who want an agent that briefs them, chases work in the background and drafts for approval. The vendor describes it as: Claude Duet is Anthropic’s proactive agent: you hand off work by text or voice message, it keeps working on its own cloud computer while you are away, and it comes back with finished docs, sheets, decks or dashboards - or asks when it needs your approval. It shares your Claude account’s memory, skills and 6,000+ connectors.
Who should look elsewhere?
Anyone outside the US, on a free or Team/Enterprise plan, or needing errands, bookings or live browser transparency - Anthropic says Duet is not built for those.
How much does Claude Duet cost?
Requires Claude Pro ($20/month, $17/month billed annually) or Max (from $100/month) - read 26 October 2026. Duet itself has no separate price; research preview, waitlist, US only.
What can it access and what needs my approval?
Approval-first for sensitive actions: before sending a message, paying or deleting, it shows the action, destination and content including every recipient, and waits (vendor safety page, read 26 October 2026). Some actions can never skip approval (paying or sending money, deleting an order, creating an account); securities trading, inter-account transfers and accessing others’ accounts are off-limits entirely. Purchases use one-time merchant-and-amount-locked cards from a wallet such as Link from Stripe, approved in the wallet app; the real card number is never shared with Duet or the merchant. Passwords, cookies and sign-in tokens never reach the model; they pass through separate machines the model cannot access.
What happens to my data?
Vendor safety page (read 26 October 2026): approvals and safety checks run outside the model on separate machines; an outbound-message classifier screens drafts, with “never share” treatment for credentials, card and ID numbers, and consent-based sharing for health or finance details. Conversation history and memory persist until the user deletes them; deleting the account wipes everything Duet stored. Anthropic cites external penetration testing and red-teaming of credential features. No statement on model training was found on the Duet safety page; the general Claude privacy terms apply.
Has AgentComparison tested it?
No. This record is built from the vendor's own documentation, checked 2026-10-26. Vendor claims are attributed, not confirmed.
How to evaluate Claude Duet yourself
- Pick one low-risk task that matches its documented capabilities: proactive morning brief and attention-flagging across connected inbox and calendar.
- Write the expected output and what counts as failure before you start.
- Connect the minimum access needed, and check approval settings first.
- Run it twice and compare, then review every action it took.
- Check cost after any free allowance. Unknown price is not free.
Related
Sources and history
- https://claude.com/product/claude-duet · official · read 2026-10-26
- https://claude.com/resources/articles/claude-duet-safety-privacy-and-security · official · read 2026-10-26
- https://support.claude.com/en/articles/17631495-get-started-with-claude-duet · official · read 2026-10-26
- https://claude.com/pricing · official · read 2026-10-26
Change history
- 2026-10-26: Initial documentation research from the product page, safety/privacy/security page, help centre and pricing page. Launched 22 October 2026 as a US research preview. No hands-on testing.
- 2026-11-01: Product page re-read: availability now lists Pro, Max and Ultra plans in the US (was Pro and Max); powered by Opus 5.5 noted. Waitlist and research-preview status unchanged. checked_label stays 2026-10-26 (partial re-check only).
Documentation overview expanded with current pricing/access sources. No hands-on results claimed.
Report a correction