Claude Duet Explained: Access, Background Work and Shared Usage Limits
What changed
Claude Duet is Anthropic’s new proactive assistant preview. Rather than waiting for every prompt, it can check connected apps, prepare work and return when a task needs attention. Anthropic announced it on October 21, 2026. The useful distinction is between asking for an answer now and handing off a job that should continue while you are away. That difference matters more than the name of the app.
This guide checks the public announcement and documentation as of October 27, 2026. It is a documentation-based explanation, not a claim that we have tested Duet’s full workflow or measured its reliability.
Who can get access, and what does it cost?
The current preview is for US Claude Pro and Max subscribers with a waitlist invitation. A paid subscription alone does not guarantee admission. Team and Enterprise are not included yet. During the preview, Duet has no separate subscription charge, but uses the same Claude plan allowances. The invitation setup is tied to the account used to join. Check the official instructions if an invitation and the account on your phone do not match.
Anthropic describes standalone iOS and Android apps and access through claude.ai. Availability is a rollout condition, not a reason to buy an account from a stranger or assume a promised invitation date.
What happens while you are away?
Anthropic says Duet uses a cloud computer and browser, so work can continue when your own device is off. It prepares a morning brief and checks connected accounts periodically. These checks are not continuous real-time monitoring. Its design calls for review before messages, purchases, signatures and deletions. Those are product descriptions, not a guarantee that every decision will be correct.
A useful way to frame a job is to specify its deliverable and stopping point. “Find three venue options with current opening hours and save the links; stop before booking” is an original example. The output is concrete enough to inspect, and a booking is a separate decision. A vague instruction such as “sort everything out” gives you no clear way to judge completion.
Choose a first task you can actually check
The following is our original task-selection worksheet, not a benchmark or a list of guaranteed Duet capabilities. Start with a small task whose inputs and output you understand. Increase the scope only after checking the result.
| Task | Useful deliverable | What you check | Where to stop |
|---|---|---|---|
| Prepare for a meeting | A short brief with source links and unresolved questions | Correct person, company and meeting date | Before contacting anyone |
| Compare three products | A requirements table with current manufacturer references | Exact model, region and price assumptions | Before purchase |
| Triage a small email set | A proposed priority list and unsent draft replies | Whether each message is understood in context | Before sending or deleting |
| Research an unfamiliar concept | A short explanation and links to original references | Whether the references actually support each claim | Before treating the answer as verified |
For a first comparison, pick products you already know rather than an urgent high-stakes decision. For a brief, choose a meeting whose details you can independently confirm. This makes the exercise a test of your own workflow: did the assistant deliver the requested artifact, preserve the constraints and flag what was missing? A polished paragraph by itself does not answer those questions.
Background work still uses your Claude allowance
Claude distinguishes usage over time from the amount that fits into one conversation. Paid plans have five-hour session and weekly usage limits; activity across Claude surfaces shares the allowance. Task complexity, conversation size, model and effort can affect consumption. A larger context window is not unlimited work.
Duet’s checks and background tasks also count toward the shared plan limit. Included in your subscription therefore does not mean cost-free in terms of allowance. If you mainly want an occasional explanation, compare that need with a conventional chat before turning a broad ongoing task into background work.
Here is a simple decision exercise, not a token estimate: write down which recurring check you want, what should trigger an update and what would make the check unnecessary. “Tell me when this specific reply arrives” is a clearer objective than “watch everything.” The product’s actual checking schedule and task controls still determine what it can do. We have not measured a per-task message cost and do not assign an invented daily quota.
What carries over, and what should you verify?
Duet reuses the memory, skills and app connections associated with the Claude account, according to Anthropic. Its help page currently lists no model picker, no way to watch or take over its cloud browser, no phone calls or outgoing text messages, and no Microsoft 365 support. These limitations can change during the preview.
For research, distinguish finding information from checking its evidence. A connector can supply a document; that does not make the document’s conclusions correct. Read the original reference for a claim that matters, and ask the assistant to separate what it found from what it inferred.
We tested Ithy Research Library search and fetch in ordinary Claude chat: it retrieved an Ithy diffraction explanation with links to OpenStax. That is a Claude demonstration, not a Duet compatibility test or an Anthropic endorsement. The Ithy connection guide contains the actual example. Anthropic’s custom-connector instructions also explain that remote connections run from its cloud infrastructure, rather than directly from your laptop. If a connection appears in one Claude client, confirm what tools are available in the client where you intend to use it.
A useful first prompt
Try an explicit, reviewable request such as this original example:
Find two reliable explanations of light diffraction, open them, and compare the assumptions in their worked examples. Link each source. Mark any calculation you could not verify. Prepare a short note for me to review; do not send it anywhere.
The wording sets a research question, a source requirement, an output and a boundary. Judge the result against those requirements. If it returns only a general description, ask for the missing comparison. If a link fails or a source does not support the claim, that is a limitation to resolve, not something a confident summary can hide. This is a proposed test prompt; it is not a report of a Duet test we performed.