Anthropic's Claude Duet is a waitlisted proactive agent on Opus 5.5 that asks before it sends, pays or deletes, and its monitoring draws on your normal Claude limits
Research preview opened 21 October for US Pro and Max subscribers: morning briefs, background tasks and 6,000+ connectors. Verdict: worth a low-stakes trial if you live in inboxes and Slack, but the security claims are Anthropic's own and more detail is promised in the coming days.
It comes to you with what needs your attention, completes tasks you ask of it, keeps working even when you step away, and checks with you before it acts on your behalf.
Comprehensive take
A coherent answer to the approval problem: preparing work is autonomous, consequential actions are gated, and a few actions are banned outright. The safety evidence is still vendor-stated and the usage-limit cost is unquantified, so start small and measure before connecting anything sensitive.
Use cases
Founders and small-business owners drowning in triage
Use the morning brief and background monitoring to surface what needs a decision, with sends, payments and deletions held for your approval.
Consultants and account managers
Anthropic says Duet drafts, schedules and prepares for calls from connected apps; Reuters lists meeting scheduling and financial forecasts among its tasks.
Security-minded teams evaluating agents
The credential design (no direct view of passwords or tokens), the Link spend approval and the bug-bounty coverage give you concrete controls to test.
Specs
| Model | Built around Claude Opus 5.5 source |
|---|---|
| Platforms | iOS, Android and claude.ai; waitlist source |
| Pricing | Included in Pro and Max during the preview; same usage limits as Claude source |
| Initial availability | US Pro and Max subscribers (Reuters) source |
| Connectors | Over 6,000 in the Claude Marketplace source |
Insights
- Axion reading: the bundling means the real price of Duet is the share of your Pro or Max allowance that background monitoring consumes; Anthropic has published no figure for it.
- Axion note: the penetration test, classifier and bug-bounty coverage are Anthropic's statements; no test results are public, and Anthropic says more safety detail is coming.
- Axion note: Anthropic itself says Duet can misjudge urgency and cannot yet show its browser, so the oversight you get is the approval step, not a live view.
What Anthropic actually launched
On 21 October Anthropic opened a research preview of Claude Duet, a separate app (iOS, Android and claude.ai) in which Claude behaves as a proactive agent rather than a chat window. Anthropic's pitch is that it brings you what needs attention, does the tasks you hand it, keeps working when you close the laptop, and asks before acting for you. It is aimed at professionals and small-business owners, is built around Claude Opus 5.5, and sits behind a waitlist. Reuters and TestingCatalog both report that access starts with US Pro and Max subscribers; Anthropic's own post says only that it is included in Pro and Max with the same usage limits as the rest of Claude.
The product has three visible mechanics. A short morning brief summarises the most important items across your connected apps. During the day Duet watches those apps (TestingCatalog reports checks every few hours) and messages you only when it judges something needs you. And it can take on tasks: Reuters lists scheduling meetings, financial forecasts and web apps. It inherits memory, skills and connected apps from your Claude account and can use the 6,000-plus connectors and plugins in the Claude Marketplace, per Anthropic.
The part that matters: how much it is allowed to do
The design question for any always-on agent is where the approval line sits. Per Anthropic and the reporting, Duet can read, draft and prepare on its own but must show the proposed action and wait before it sends a message, pays, signs or deletes. Securities trading and moving money between accounts are off-limits even with approval. Purchases go through Stripe's Link, where you approve the spend, and Reuters adds Shopify's Shop Pay. Anthropic says credential handling follows its Managed Agents design, so Duet never directly sees passwords or sign-in tokens, and that red teams and an external firm penetration-tested it and it is covered by the public bug bounty. A dedicated classifier and safety system prompt checks each message for leaks of sensitive information.
None of those security claims has been independently tested yet, and Anthropic says it will publish more on the safety design in the coming days. The company is also candid about limits: Duet can misread threads and rank urgency badly, is weaker at personal errands such as booking tickets, and cannot yet show you its browser, which limits how much you can see of what it is doing.
The cost nobody has priced
Duet is not billed separately in the preview, but background monitoring is not free. TestingCatalog notes that checks consume plan limits and larger projects take more model turns, so heavy users may find Duet competes with their normal Claude usage. Reuters frames the launch as a response to OpenAI's always-on "dots" agents (launched last month) and Meta's Muse agent.
Who should try it, who should wait
Join the waitlist if you live in email, Slack and calendars and lose time to triage; that is the use case Duet is built for and the one with the clearest approval model. Wait if you need non-US access, handle regulated data, or want audited evidence of the safety controls before connecting accounts. The first useful test is simple: connect one low-stakes inbox, see how often the urgency ranking is wrong, and measure what it does to your usage limits before widening access.