Anthropic announced Claude Duet on October 21, 2026, opening a research preview waitlist for U.S. Pro and Max subscribers. The agent runs on Claude Opus 5.5 — the model Anthropic describes as built for "long-running agentic coding and complex knowledge work" with a 1-million-token context window and 40% lower operating cost than its predecessor.
This is not a chat interface with a new coat of paint. Duet is a proactive agent: it monitors connected applications (email, calendar, GitHub, Notion, Linear, and more), prepares work in the background, surfaces daily briefs with what needs your attention, and — critically — keeps running tasks after you step away. The work does not pause when the device sleeps or the browser closes.
What Duet Actually Does
According to Anthropic's announcement and early preview coverage, Duet's capabilities center on three loops:
- Monitoring: Duet watches connected apps for changes — new emails, calendar updates, GitHub notifications, Linear tickets, document edits. It builds a running model of your work context.
- Preparation: Before you sit down, Duet drafts responses, summarizes threads, prepares code changes, and assembles briefing documents. It presents these as proposals requiring approval — it does not act autonomously on high-stakes actions.
- Continuation: When you assign a task ("refactor the authentication module," "prepare the Q3 board deck"), Duet persists the work in a cloud session. Closing the app or putting your laptop to sleep does not interrupt it. You return to a completed or advanced task.
Anthropic's framing is deliberate: "It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf."
The Technical Substrate: Opus 5.5 and MCP
Duet runs on Claude Opus 5.5, which ships with several capabilities relevant to persistent agents:
- 1-million-token context window — enough to hold entire codebases, long email threads, and project history simultaneously.
- Preserved thinking — Anthropic's anti-distillation safeguard (introduced with Fable 5.1) that keeps the model's reasoning traces intact across sessions, making long-horizon planning more coherent.
- 300k output tokens via the
output-300k-2026-03-24beta header on the Message Batches API — enabling genuinely large single-turn productions (full features, comprehensive reports, multi-file refactors). - 85% reduction in containment bypass attempts — a safety metric Anthropic highlighted in their September 2026 threat intelligence report, relevant for an agent that operates with broader tool access.
For developers on the MCP stack, the architectural implication is clear: Duet is a reference implementation of a persistent MCP client. It maintains long-lived sessions, manages tool discovery across dozens of connected services, and handles the session lifecycle (reconnection, state sync, budget enforcement) that every production agent runtime must solve.
What This Means for MCP Server Builders
If you are shipping an MCP server today, Duet changes the integration surface in three ways:
1. Servers must handle long-lived, stateful sessions
Duet does not open a connection, call a tool, and disconnect. It maintains persistent connections to your server across hours or days. Your server needs stable session IDs, resumable streams, and graceful handling of client reconnection with context replay. The MCP 2026-07-28 spec's stateless HTTP core helps here — but your application logic must track what Duet has done, what it's doing, and what it should do next.
2. Tool descriptions become agent interfaces
Duet discovers your tools dynamically and decides when to call them. Your tool schemas, descriptions, and error messages are now agent-facing documentation. Vague descriptions like "search the database" will produce confused agent behavior. Precise, structured schemas with clear preconditions, side effects, and idempotency guarantees become a competitive advantage.
3. The "human in the loop" is a first-class protocol concern
Duet's design mandates approval gates for high-stakes actions (sending messages, spending money, signing documents, deploying code). MCP servers that expose such actions should consider returning proposal objects rather than executing immediately — letting the client (Duet, or any well-behaved agent) surface the proposal for human review. This pattern does not exist in the MCP spec today; it is an emerging convention that Duet will pressure-test.
The Competitive Landscape
Duet enters a category Anthropic is effectively defining: always-on personal agents. The nearest analogs:
- OpenAI's Operator (research preview, browser-based, task-focused, requires active supervision)
- Meta's AI agents (integrated into WhatsApp/Instagram, social-first, less developer-facing)
- Microsoft Copilot (deep OS integration, enterprise-focused, reactive)
Duet's differentiation is the background persistence combined with proactive monitoring. Operator waits for a task. Duet watches your inbox and calendar and proposes tasks. For developers, this means the "agent entry point" shifts from "user types a prompt" to "agent observes a state change and prepares a response."
What to Watch
The research preview is limited: U.S. only, Pro/Max only, waitlisted, and explicitly labeled "early." Three signals will tell us whether Duet becomes a platform primitive or a niche product:
- MCP ecosystem adoption: Will major MCP servers (GitHub, Linear, Notion, Google Workspace) add Duet-specific affordances — proposal returns, webhook-style push notifications, incremental sync endpoints?
- Opus 5.5 API availability: Duet's capabilities depend on the model. If Opus 5.5's 300k output tokens and preserved thinking remain beta-gated or Pro/Max-exclusive, the developer story stalls.
- Session economics: Persistent agents consume tokens continuously. Anthropic has not published Duet's pricing model. If a day of background monitoring costs more than the human time it saves, adoption will be hobbyist-only.
The Takeaway for Builders
Claude Duet is the clearest signal yet that the agent runtime layer is moving from "chat with tools" to "persistent background process with human oversight." The MCP protocol gives us the wire format. Opus 5.5 gives us the model capacity. Duet shows the product shape.
If you are building MCP servers, start designing for long-lived sessions, proposal-based tool execution, and agent-discoverable capability surfaces. If you are building agent applications, Duet's architecture — monitor, prepare, propose, persist — is a template worth studying.
The waitlist is open at claude.ai/duet for eligible U.S. subscribers. The rest of us watch, wait, and build the infrastructure the next wave will run on.
If this was useful, you can support my open-source work on Ko-fi or check out my services.