📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
Summary: On November 5, the OpenClaw Foundation shipped Multiplayer: two laptops driving one agent, sharing one context, with every request routed through a single queue. What it deletes is the most expensive step in collaboration, explaining again what you already did. KAIHE AIBOX(铠盒AIBOX), a dedicated agent computer, happens to be the half this setup usually lacks: a machine that stays on and does not tie up your laptop.
Two minutes at the booth in San Francisco
Two laptops, one agent, a large display rendering Fort Mason in real time. Visitors picked a landmark from somewhere else; the Statue of Liberty won most often.
Someone typed Add the Statue of Liberty into the first laptop. Before they hit enter, the same sentence was appearing letter by letter on the second screen. The statue landed in front of the building, lumpy and low-poly, but recognisable. Then someone walked to the second laptop and typed Make that three times bigger, and the statue tripled. Nobody explained which object "that" meant, and nobody asked the agent to reconstruct a colleague's work.
The problem the publisher admits itself
The blog is blunt. Every department has its own agent and its own model. Real collaboration starts when someone has to explain what they asked their agent to do, what it produced, and how they got there, and the colleague then feeds an approximate version of that into their own agent, hoping the important parts survive.
Four short lines: context gets lost, instructions are mangled, changes collide, the pact to coordinate breaks. Multiplayer removes that step by putting two people in one workspace, with one agent and one context.

The shared queue handles collisions
The most concrete line in the documentation: requests go through a shared queue, so you can see the agent is busy instead of blindly throwing conflicting instructions at it.
Three layers sit on top. Every session records an immutable creator, an assignable owner, and the history of who actually prompted. Presence shows who is viewing and who is typing, with drafts that never reach the model or the transcript.
Guardrails, not isolation
The documentation is clearer than the announcement. A gateway is a single trust domain: anyone who can message a tool-enabled agent shares that agent's tool authority. Session ownership and presence are usability features, not isolation between people who distrust each other. For mutually untrusted groups, the guidance is one gateway per group, ideally on separate accounts or hosts.
It wants a machine that stays on
The prerequisites are short: a host that stays on, a small VPS, an office Mac, or any supported install target. The two laptops in the demo are just input surfaces. The machine actually running the agent cannot shut down, cannot sleep, and cannot leave with whoever goes home.
| The team | One shared agent | Everyone separate |
|---|---|---|
| Two or three who trust each other | works, skips the retelling | works, synced by hand |
| Five or six needing ownership | works, with owner and presence | context drifts |
| Groups that distrust each other | wrong fit, one gateway each | separate deployments |

The machine underneath
In practice it looks ordinary. Tasks are handed out in the team chat during the day; two laptops pick up the same agent and edit the same work. You close the lid and go home, and the machine keeps running. Credentials and files stay on the device, permissions are bounded per person, and the hardware plus agent system is bought once with no recurring software subscription.
The model bill is still yours. Worth saying plainly: Multiplayer is free software.
Three things it does not fix
Free is not the same as costless. You bring your own model, OpenAI, Qwen, Claude or something local, and you pay for tokens.
The blog jokes that it will not cure bad management, nor the person who schedules a meeting about scheduling meetings. Anything stuck in process stays stuck; the tool only catches the technical half.
And one person does not need it. If your team is you, Multiplayer is currently a queue with decent manners.
Related Reading: - How Should Multiple AI Agents Split the Work? - OpenClaw or Hermes on a KAIHE AIBOX? - Is OpenClaw Safe? Three Layers and Six Principles
KAIHE AIBOX #OpenClaw #Multiplayer #AI agents #agent computer #local AI
For more information, search [KAIHE AIBOX] or contact: [email protected]
KAIHE AIBOX · 7x24 Personal AI Assistant | OpenClaw