A client follow-up looks simple until the email depends on a deal stage in Salesforce, a promise made in Gmail, and a decision buried in Slack. This Claude Duet review asks whether Anthropic’s new proactive agent can bring those facts together before it suggests what to send.
This review is based on Anthropic’s launch material and early Reuters and CNBC coverage. We have not tested the preview.
Key takeaways
- Claude Duet is a waitlisted research preview for professionals and small-business owners. It watches connected apps, brings work to the user’s attention, and can continue tasks while the user is away.
- Its strongest documented use case is preparing and prioritizing work across a busy day. Anthropic says it asks for approval before acting on a user’s behalf.
- Anthropic acknowledges that Duet can misread a thread or misjudge urgency. For a client follow-up, the practical test is whether it has the latest deal status, correspondence, and internal decision.
- A permissioned, source-cited company brain could supply that context if Duet can accept an external MCP server or a comparable connector. The launch sources do not establish that connection.
What is Claude Duet?
Claude Duet is Anthropic’s proactive AI agent for professionals and small-business owners. Rather than waiting for every prompt, it checks connected apps, prepares a daily brief, and surfaces work that may need attention. It can carry out requested tasks while the user is away, then seek approval before acting on their behalf. It launched as a waitlisted research preview on October 21, 2026.
Anthropic says Duet is available through a standalone iOS or Android app and claude.ai. During the preview, access is included with Pro and Max plans and draws on the subscriber’s existing Claude usage limits. Reuters reports that initial access is for US subscribers and that Duet is built around Claude Opus 5.5.
This is an early-access product, so its documented capabilities are a better basis for evaluation than assumptions about how reliably it will perform on any particular company’s work.
What does Claude Duet do well?
Duet’s useful idea is to prepare work before someone remembers to ask. Anthropic describes an agent that flags matters needing attention, keeps working after the user steps away, and checks before taking action on the user’s behalf. For an owner moving between sales, scheduling, and delivery, that could make a daily brief more useful than another empty chat window.
The range of tasks is substantive. CNBC reports that Duet can build dashboards, schedule meetings, and create financial forecasts. Reuters also lists web apps among the tasks Anthropic says it can handle. Those are reported capabilities, not results from our own testing.
Duet also has a meaningful route into existing work. Reuters reports that it can use more than 6,000 connectors in the Claude Marketplace, including Gmail and Slack. That matters for follow-ups: an agent connected to the relevant inbox and conversation has a better starting point than one working from a pasted summary.
Finally, the approval step addresses a real distinction between drafting and sending. A suggested client email is easy to revise. An email sent with the wrong commitment is harder to take back. Anthropic says Duet checks with the user before it acts on their behalf; teams should still verify the exact approval behavior for the actions they plan to allow.
Claude Duet review: where does it fall short?
The clearest limitation is one Anthropic has already named: Duet might misread a thread or miscalibrate what is urgent. A prompt to follow up with a client can contain several plausible next steps. The right one may depend on an internal decision made after the last email.
Connected apps alone do not prove that an agent has the complete, current account picture. This is an evaluation risk, not a claim that Duet lacks access to a particular app. For example, access to a Gmail thread would not, by itself, establish whether the CRM stage changed yesterday or whether a private Slack discussion ruled out offering a discount. The launch material does not show how Duet reconciles conflicting account facts or cites them in a proposed follow-up.
That gap becomes more important because Duet is proactive. If it surfaces a follow-up at the wrong moment, the user has to inspect the evidence before deciding whether the suggestion is useful. The agent’s judgment and the quality of its available context both affect that decision.
There are also limits outside sales work. CNBC reports Anthropic’s assessment that Duet is less strong at personal errands such as booking tickets. Its stated focus is workplace assistance, which is the fairer standard for this review.
How would Claude Duet improve on a company brain?
A company brain is a permissioned, source-cited knowledge base drawn from the company apps a team chooses to index. Gyld makes that knowledge available to AI agents as an MCP server. For Duet, the potential benefit would be a proposed action grounded in the latest relevant account facts, with links a person can check before approving it.
Consider a fictional client, Acme. The owner asks Duet to prepare a follow-up after a product call. These are illustrations of what additional context would change, not accounts of Duet or Gyld completing the tasks.
| Task | Without company context | With a company brain |
|---|---|---|
| Draft the Acme follow-up | The draft may rely on the last visible email and ask for a decision the client has already made. | The agent checks the latest permitted Gmail exchange and Salesforce deal record, then drafts against the current stage with source links for review. |
| Address a requested discount | A polite reply might offer terms the team has not approved. | A permitted Slack decision and the signed terms in Google Drive show what can be offered; the draft cites those sources and leaves the decision with the owner. |
| Decide whether to follow up today | A quiet inbox can make the account look stalled. | A current CRM update and an internal handoff note show who is waiting on whom, so the suggested next step matches the account’s status. |
Permissions matter as much as retrieval. A private pricing discussion should not become company-wide context simply because it is relevant to a deal. Gyld’s approach to a managed company brain lets a team choose what to index and permission knowledge as private, team, or company-wide. Source citations then give the reviewer a way to check the facts behind a draft. Teams comparing that approach with a pipeline they maintain themselves can examine Gyld’s comparison with other context approaches.
There is an integration condition here. Gyld exposes company knowledge through MCP servers, but the supplied Duet launch sources do not establish whether Duet accepts an external MCP server. If it does, that would be the route to providing this context; otherwise, Duet would need MCP support or a similar connector first. Duet does not have a documented Gyld integration in these sources.
A useful pilot would use one account and one decision: connect only the permitted sources, ask for a follow-up draft, and check its deal stage, promises, citations, and proposed recipient before anyone sends it. That tests whether the context actually improves the work.
What would a company brain not fix?
Better account evidence would not guarantee sound judgment. Duet could still misread a nuanced exchange or rank an ordinary message above an urgent one, the risks Anthropic acknowledges. Citations make a mistake easier to catch; they do not turn a draft into an approved business decision.
Context also cannot change preview access or usage limits. Anthropic says users need to join a waitlist and that Duet shares the usage limits of their Pro or Max plan. Nor would company records make a workplace-focused agent better at the personal errands Anthropic says are less strong.
Verdict
Claude Duet is worth watching for a professional or small-business owner who wants an agent to notice work, prepare a brief, and keep requested tasks moving. The documented app access and approval step make it especially interesting for preparing work that a person will review.
Wait before relying on it to send or prioritize consequential client follow-ups. The preview needs to show, in your own account, that it can identify the current deal state, respect access boundaries, and make the basis for its suggestions inspectable. This Claude Duet review would become more favorable for that use case if Duet documents a path to external permissioned context and demonstrates reliable, source-backed drafts. Until then, keep a person responsible for the final message.
Frequently asked questions
Is Claude Duet available now?
Claude Duet opened as a waitlisted research preview on October 21, 2026. Anthropic says eligible Pro and Max subscribers can use it through iOS, Android, or claude.ai; Reuters reports that initial access is limited to US subscribers.
Does Claude Duet connect to Gmail and Slack?
Yes. Reuters reports that Gmail and Slack are among the Claude Marketplace connectors Duet can use. Whether those connections provide every fact needed for a particular task depends on what the user has connected and what information is available there.
Can Claude Duet send a client email without approval?
Anthropic says Duet checks with users before acting on their behalf, and Reuters reports that sensitive actions require approval. Check the preview’s controls for your exact email workflow before granting access or relying on that safeguard.
Does Claude Duet support external MCP servers?
The launch sources supplied for this review do not establish whether Duet accepts external MCP servers. Duet’s reported access to Claude Marketplace connectors should not be taken as confirmation that a third-party MCP server can be added to this preview.
What should a small business test first?
Test a draft that crosses sources, such as a follow-up that needs the latest email, deal status, and approved terms. Review the proposed facts against those records, confirm who can see each source, and require approval before sending. For a narrower starting point, see how to pick a first AI workflow for a small business.
Related reading
- Personal AI Agents for Work: The Company Context Test offers a way to evaluate whether an agent knows enough to act on work information.
- AI Agent Action Permissions: Who Can Submit a Form? examines the approval boundary between preparing an action and taking it.
If you want your existing AI tools to work from permissioned, source-cited company knowledge, start building your company brain with Gyld.