Claude Duet: Anthropic's Cloud Agent Turns Your Phone Into a Remote Control
Cloud-hosted agents that operate independently of the phone will become a new class of client for web services and native apps. Developers have to assume an AI is on the other end of their login flows, payment forms, and data pipelines, and they will need agent-friendly entry points, not just human-oriented UIs. The explicit ban on trading and bank transfers signals where the trust boundaries are drawn, but the larger shift is that apps must now be built to be used by agents at scale.
Claude Duet runs on Opus 5.5 in an Anthropic-managed cloud VM and an isolated browser sandbox. The phone app does nothing but accept text or voice input, surface approval prompts, and relay messages. All real work, reading email, filling forms, calling APIs, browsing the web, happens remotely, so closing the phone does not stop the agent. It shares account memory and usage quotas with the main Claude app, and it reaches 6,000+ third-party services through the Claude Marketplace.
The capability list is intentionally bounded: scheduling meetings, drafting emails, checking unread messages, handling invoice collections, running payroll via Gusto, building dashboards, financial forecasting, and prototyping web apps are allowed. Sending messages, paying with a Stripe Link virtual card, signing with DocuSign, and deleting content require explicit approval. Securities trading and bank transfers are banned outright, even with user consent. The agent also cannot book flights or restaurants, and it cannot show its browser screen to the user.
Architecturally, the phone sits in an entry layer that never reads the screen, uses accessibility services, or controls other apps. That sets Duet apart from Google's Gemini Intelligence, which operates on-device on Pixel and Galaxy hardware and uses system-level automation over native apps. Duet's cloud-first route also differs from OpenAI Dots and Meta Muse. All three push from chatbot to autonomous doer, but Duet is the one that keeps the phone a pure remote control.
Anthropic deliberately capped the agent's reach on the phone: no screen reading, no accessibility service, no control of other apps. This keeps the trust surface small but also makes the phone a pure I/O device, a design choice that avoids the privacy friction of Gemini Intelligence's on-device automation.
The shared memory and quota between Duet and the main Claude app treats the agent as an execution arm of the same account, not a separate product, which could push users to keep personal and work contexts in a single AI system.
By running all execution in a cloud sandbox with 6000+ connectors, Duet makes the browser and web APIs the primary interface for agent work. That will pressure web products to authenticate and handle agent traffic without a human at the wheel.