How to Use Claude Duet: Setup, Quota and First Week

Claude Duet works best with standing triage and drafting instructions, but background checks share your standard chat quota. Sloppy setup burns limits fast.

Claude Duet is Anthropic's proactive agent layer on top of Claude, launched as a research preview on October 21, 2026. Regular Claude waits for a prompt; Claude Duet keeps one continuing thread, checks connected apps every few hours, posts a morning brief, and keeps working in its own cloud computer after you close the app. It runs on Claude Opus 5.5 with no model picker, uses your existing memory, skills, connectors and searchable chats, and asks for approval before sending, paying, signing, or deleting anything.

Get in first: waitlist, eligibility, and the email trap

Claude Duet access is gated by three conditions at once: a US Claude account on Pro, Max, or Ultra, plus an invitation from the waitlist. Team and Enterprise accounts are not eligible yet, and Anthropic says those plans will come later. There is no separate Claude Duet fee during the preview. Access rides on the plan you already pay for.

Waitlist invites roll out in waves: US Max first, then US Pro, with international access pending (@ahuja_karan). Non-US accounts wait under current rules.

The practical failure point is email matching. Use the exact email tied to your Claude account when you join the waitlist, and open the invite on a phone whose Apple or Google account is linked to that same email. The Claude help center lists mismatched emails as one of the four standard reasons people think they were skipped. The other three are waves, US-only scope, and Team/Enterprise plans.

One quota detail matters before you install: Anthropic is issuing full limit resets for the preview through Settings > Usage on web or Claude Desktop. Pro accounts get 3 resets, Max and Ultra get 1 each, and unused resets expire on December 15, 2026. Both the five-hour session limit and the weekly limit refill. The resets exist because Claude Duet background work counts against the same limits at the same rate as normal chats.

Connect only what you will actually let it touch

Claude Duet inherits the connectors already on your Claude account, and onboarding surfaces the ones it already knows. Do not connect everything on day one. Connect the inbox and calendar you triage daily, then add one system of record at a time.

Named connectors and limits, per Anthropic docs and the Claude Duet product page:

Available examples: Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, DocuSign, Calendly, Stripe, plus Link by Stripe and Shop Pay for approved purchases inside Duet.
Directory examples: more than 6,000 connectors and plugins, including Shopify, Gusto, Notion, Canva, and Ramp.
Unavailable or limited: Microsoft 365 and Personal Finance are not available yet; Slack has reduced functionality.

Authentication uses OAuth, and Anthropic says passwords and sign-in tokens stay out of the chat. When the cloud browser hits a password, you type into a separate form on your device that feeds the browser directly. You can disconnect any app at any time, which stops future access, but information already stored in memory may persist. Past-chat search covers Claude Duet chats and Claude chats outside projects, follows the existing Search and reference chats setting, and deleting a chat removes it from search.

Because email bodies and web pages can contain prompt injections and no defense is perfect, Duet gates consequential actions behind manual approval.

Purchases have an extra guardrail worth knowing: Link and Shop Pay run on one-time cards scoped to the single purchase you approve, and both wallets currently live only inside Claude Duet rather than in the general Claude connectors directory.

Give it a job description, not a prompt

Claude Duet works best from standing instructions in one thread, not from clever one-off prompts. Keep a main thread for continuity, then open separate channels for topics such as inbox, hiring, or a project. Memory is shared across channels, so a preference you set once carries over.

Setup takes minutes if your Claude context is healthy:

Sign in with your Claude account and let memory, skills, projects context, and connectors carry over.
Connect Gmail, Calendar, and one source of truth such as Drive or Slack first.
Turn on notifications so completions and approval requests reach you.
Write one standing brief: what to watch, how often to report, and what always needs approval.
Assign the first small job by text or voice and check the Tasks tab, not just the chat.

Five starters copied from Anthropic's own examples, tightened so they include an approval boundary:

"Every morning, tell me which emails need my attention and draft replies for me to approve. Do not send anything."
"Clear out anything in my inbox that I don't need to read, and list what you archived."
"Check Stripe for unpaid invoices and draft reminder emails for me to approve."
"Pull together a brief on this company before my 2 pm meeting, with sources for every number."
"Watch this thread and follow up if it goes quiet for two days. Draft the nudge, don't send it."

Voice works for the same jobs. One early user dictated lead research, market research, and a monitoring task in voice mode while driving to a client meeting, with sub-agents fanning out the work in the background (@ZondarkAI).

What matters after day one is quota. Zapier's hands-on test called the unsolicited security-alert summary impressive, then burned through Pro session limits in the first couple of days because periodic checks kept billing quota. Decide up front which topics get proactive monitoring and which stay on-demand.

Run the first week without burning quota

The Tasks tab is your control room. It shows what Claude Duet is running and what needs your input, while the main conversation stays quiet except for completions and decisions. Only you can mark a task complete, even if Duet suggests it is done. Finished artifacts such as research briefs, financial models, slide decks, and dashboards land in the Make tab.

Because background check-ins and multi-step jobs bill your normal session and weekly limits, one tester complaint keeps recurring:

"kicking off new sessions per thread and scheduled ask creates huge token overhead ... overly asks for permissions ... needs to be able to login to google" — @opensesahme

Three guardrails keep the burn under control:

Keep monitoring intervals loose. Anthropic checks connected accounts every few hours, so time-critical alerts belong outside Duet.
Batch related work and approvals instead of drip-feeding single-message tasks, because each new session adds overhead.
Watch Settings > Usage after enabling recurring checks, and keep a preview reset in reserve. Anthropic notes that large jobs such as financial models and decks cost more turns.

While waiting for you to approve a consequential action, Duet continues non-blocked background work and returns later. Hard limits are in the table below.

Expect judgment errors. Anthropic, Zapier testers, and the unofficial guide at claudeduets.com all warn about the same trio: misread threads, misjudged urgency, and gaps between periodic checks. Review every draft and approval request before confirming, and keep time-critical alerts outside Duet for now.

What to hand it versus what to keep

Hand Claude Duet work where a draft is the deliverable and a human approves the send. Keep work where timing, money movement, or local access is the constraint.

Hand to Duet	Keep outside Duet
Morning brief over inbox + calendar with drafted replies	Time-critical paging between check-ins
Unpaid-invoice scan + reminder drafts via Stripe/QuickBooks	Securities trading or account-to-account transfers
Meeting briefs with sourced numbers, decks, dashboards	Phone calls, SMS, local files, full Claude projects access
Hiring-panel note collection and offer-email drafts	Microsoft 365 workflows until the connector ships
Renewal, thread, and deadline watch with nudges held for approval	Anything needing a visible browser takeover or captcha handling

Duet ships as standalone iOS, Android, and web at claude.ai, with no desktop client; a Mac app is in the works (@ahuja_karan). It can read cloud conversations but cannot interact with those threads or local Code sessions without per-message approvals (@LLMJunky; @Rakel076; @codewithpassion).

Duet suits approval-gated triage across connected work apps. It is a weaker fit if you need desktop access or low-friction autonomous execution.

Join the waitlist if fragmented triage is your bottleneck and US-only, periodic, approval-gated operation is acceptable. Otherwise wait for global access and harder reliability evidence.