Claude Duet Is Still on the Waitlist, While OpenDuet Is Already Open Source

Anthropic's Claude Duet is still in research preview, and even Pro/Max users in the US have to queue for the waitlist. CopilotKit went ahead and open-sourced a similar product: OpenDuet.

OpenDuet is a proactive agent that doesn't wait for you to ask. In the morning it sends you a brief first; during the day it digs through email, calendar, and Slack in the background, and only surfaces when something needs a decision. Give it a task, such as scheduling a meeting or building a dashboard, and it keeps running on its own computer, then comes back when it's done. Before sending, paying, signing, or deleting, it shows you the exact action and waits for you to choose Allow once, Always allow, or Refuse.

Anthropic's Claude Duet, released a week ago, follows the same approach. It is based on Opus 5.5, proactively tells you what needs attention, can complete tasks in the background, and asks for consent before critical operations. Anthropic emphasized several safety designs: it never directly sees your passwords, a dedicated classifier blocks sensitive information, and important decisions are left to you. But it is currently available only in the US, and Pro/Max users are still on the waitlist.

OpenDuet's difference is open source and self-hosting:

• You own the agent, rather than renting it. Data stays on your own server.
• It is not tied to a model or framework. Any agent harness that uses AG-UI (an open protocol connecting front ends and agent backends) can plug in, including LangGraph, Mastra, CrewAI, and Pydantic AI. The default model is OpenAI's gpt-5.6-terra, and you can also switch to Claude Sonnet.
• Approvals are core. All operations such as sending, paying, signing, and deleting first go through a policy gateway, leave an audit trail, and then a decision is made whether to execute.
• Passwords never enter the conversation. Login credentials are entered through masked fields and stored encrypted, and are invisible in the agent's context.
• Each agent has its own computer. Chromium, /workspace, shell—even if you close your laptop, it keeps running.
• Multi-channel. It works on Web, Mobile, Slack, and Microsoft Teams.

The technical foundation is CopilotKit's OpenBot. All OpenDuet customizations are in one folder: name, Bots, channels, models, skills. Copy it, make it your own, and you have a new agent product. The repository is at github.com/CopilotKit/OpenDuet; the official description says clone this template and customize it however you want, enterprise-ready, MIT license.

It is still in early development, but core features already work: morning brief, email/calendar triage, task panel, Make (generates pages, docs, dashboards), scheduled tasks, and MCP connectors (Google Drive, Notion, etc.). It requires a free CopilotKit Intelligence project to store threads and memory.

One thing to note: the name OpenDuet is not unique. Arindam Majumder also open-sourced a project with the same name on October 23 (github.com/Studio1-OSS/OpenDuet), using Mastra, Nebius Token Factory, and Composio, under the MIT license, running open-source models. The two projects are independent and have no relationship with Anthropic.

Atai Barkai of CopilotKit said in a reply that every personal AI coworker product is essentially a different UX variant of the same core functionality; what they are building is the UX layer. That is also where OpenDuet's value lies: open-sourcing the interface and infrastructure for a proactive agent while handing approvals, data boundaries, and model choice back to users.

Some commenters said self-hosting and human approval make the phrase "your agent" feel real for the first time. Others asked how many of these agents people will actually keep using for more than a week, rather than becoming yet another dashboard you have to tend. There is no answer to that question yet.