Glove

Rebuilds

In 2026 four companies shipped agents that keep working when you are not looking: xAI's Grok Bot, Anthropic's Claude Duet, OpenAI's Dots and Meta's Muse. Each one is built here from Foundry primitives, step by step, and deployed on Kuiper. Read them to see how the parts fit together in a real product, or copy one as a starting point.

›

These are independent examples built from public docs and launch coverage. They are not affiliated with or endorsed by xAI, Anthropic, OpenAI or Meta, and they are not how those products are built. Each page lists its sources.

The same parts, arranged four ways. Grok Bot spreads work across sibling bots a person makes as data. Claude Duet keeps one partner with its own clock. A dot holds its person's responsibilities and checks each action against their Custom Rules. Muse does the work in a browser of its own, behind a Sentinel that decides what leaves it.

The four rebuilds

xAIGrok Bot

Sibling bots a person makes and names, all instances of one definition, working in the apps the person connects (Gmail, Linear). Each runs on its own requests: a message, an @-mention, a sibling's handoff or its own routine, and the starred Primary Bot takes asks that name no bot. They share group chats, one computer and a skill library.

examples/foundry-bot-crew AnthropicClaude Duet

One partner in one thread, working in the person's Gmail and Calendar. A morning brief and quiet check-ins on its own clock, new mail as it lands, approvals before it acts, and a guard on everything it sends.

examples/foundry-proactive-partner OpenAIOpenAI Dots

A dot you name, holding your ongoing responsibilities and working from a connected mailbox. Read-only research between conversations, follow-ups you ask for that act under your Custom Rules, an auto-review of sensitive actions, a pause, its own computer, and the same dot in the app, Slack or a call.

examples/foundry-standing-agents MetaMeta Muse

One personal agent per person, named and styled as they like, in the app and on WhatsApp, that does the work across the apps they connect at the access they choose. Its own browser, a separate Sentinel on everything that browser sends, logins it never sees, one-time cards, plans it keeps, and a log of what it did and plans.

examples/foundry-personal-agent

Side by side

	Grok Bot	Claude Duet	OpenAI Dots	Meta Muse
Who you talk to	The bot whose job it is, or your starred Primary Bot; bots are siblings	One partner	Your dot, in the app, Slack or a call	Your Muse, in the app or on WhatsApp
Instances	Every bot a person makes, all of one definition	A partner and a computer per person	A dot and a builder per person; specialist dots per organisation	One Muse per person
Conversations	One chat per bot; a side of each group chat; one per mesh thread	Exactly one	One per channel, sharing memory	App and WhatsApp, sharing memory
What wakes it	A message, an @-mention, a sibling's handoff, its own routine (a schedule or new mail)	You, new mail, the morning brief, check-ins	You, new mail, a Slack message, the research schedule, follow-ups and schedules you asked for	You, new mail, checks it set itself, a daily look for things to suggest
Connected apps	Gmail and Linear, per person	Gmail and Google Calendar	Gmail (the support mailbox)	Gmail and Calendar, read or act, as the person chooses
Who does long work	The sibling whose job fits, over the mesh; the Primary Bot hands specialists their part	Its computer, a durable subagent	The builder, a durable subagent	Muse, in its own browser on Kuiper
What waits for you	Sends, new skills, new apps and sign-ins	Send, pay, sign, delete	Whatever your Custom Rules say to ask about, plus deletions, installs and saved-card purchases; password changes and money transfers are yours to do	Send, buy, and any site the Sentinel asks about
The safety idea	Approval rules, and require-approval rules win; isolation is per person, so bots are not a boundary from each other	A guard on everything outbound	Read-only research, an auto-review you can't switch off, and a monitor that pauses the dot	A Sentinel it can't talk round, on everything its browser sends, with read and write granted separately

What they share

Under the differences, all four use the same handful of Foundry parts:

Agents as data. A definition says what a kind of agent can be; each agent, its name, job and preferences are an instance in its person's workspace. See Agents are data.
Other agents for long work: Grok Bot's siblings message each other on the hosted mesh by instance id; Duet's computer and the dot's builder are durable subagents their agent owns. Results come back on their own. See Multi-agent systems.
Schedules that start runs nobody asked for. See Playbooks, schedules & sleep.
Approvals on any tool that changes something outside. See Approvals & steering.
Connected apps, the same in all four. Each provider is a Foundry app installed with the person's account: its MCP server for reads, which the agent calls as functions in code; a transmission for sends, delivered by Foundry after approval; and a connection that brings new mail into a playbook. Accounts and tokens stay in your own store. See Apps & transmissions.
Transmissions and playbooks for messages from outside, such as Gmail or Slack. See Playbooks, schedules & sleep.
Kuiper for the disk the state lives on, the model gateway, and the lock on the API. See Foundry on Kuiper.

Run them

Each example runs offline with pnpm verify: the agents use short scripts instead of models, so the check needs no keys and tests the wiring end to end. Add ANTHROPIC_API_KEY and run pnpm dev to use real models in the inspector. Each README has the Kuiper commands.