Four weeks after rabbit put OS3 into general release, the number of computers one account can pair has gone from five to 20. Jesse Lyu’s team shipped OS3 on September 22 as a cloud agentic operating system that starts in chat, takes an outcome instead of a menu choice, and reaches the web, your connected apps and the machines you choose to pair. The October 10 update raised that device cap, folded the local agent into a desktop app, and claimed OS3 finished the same tasks nearly five times faster in rabbit’s first internal benchmarks while using fewer tokens.
A Month In, the Cap Went From Five Machines to 20
The OS lives in the cloud, and real computers join through rabbit agent, a small local program that turns a Windows, macOS or Linux machine into what rabbit calls a node. “We’ve spent two and a half years building toward a system you don’t operate, but instead just tell it the outcome,” Lyu said at launch, and OS3 decides whether a job stays on one node or spreads across several. The launch post capped an account at five devices, and rabbit’s support page on pairing now allows up to 20 registered computers, with a default node for work you have not assigned.
Most work runs through the node’s terminal against files and tools already there. Since the October 10 update OS3 can also open JavaScript-heavy pages in a fresh browser on the node, and when a task needs a graphical app it drives the screen with rabbit’s DLAM after you approve. A node has to stay awake and online, because a running task stops when the machine sleeps or drops off the network.
The Desktop App Installs the Agent for You
Pairing used to mean settings, register, then pasting a PowerShell or shell command with an expiring token onto the target machine. That flow still works, and it remains the only option on Linux and cloud machines, but the new macOS and Windows desktop app installs rabbit agent and connects the computer as soon as you sign in, unless you switch automatic connection off in the app’s menu. On a Mac the agent still needs Accessibility and Screen Recording permissions before it can see and drive the screen.
The October 21 update added a permission center for setting how much autonomy agents get, plus a view of background workers and their progress. rabbit’s FAQ says sensitive or irreversible actions ask for confirmation and that you can withdraw a permission at any time, and our team at TechForward has not tested where OS3 draws that line.
Your Claude Code Skills Can Move In
When you connect a computer, OS3 now looks for coding agents already installed there, including Claude Code, Codex, Gemini CLI and Cursor, and offers to import their skills and standing instructions. rabbit’s OS3 FAQ says the import reads no credentials and takes no transcripts, and that compatibility for other third-party skills varies. It is the same modular idea we covered when harnesses started loading SKILL.md folders on demand, pointed at a cloud OS instead of one coding tool.
The model is yours to bring. OS3 is free, you pay the provider whose key you add, and the FAQ names 11 supported providers plus any OpenAI-compatible endpoint or a model served from your own node through Ollama or LM Studio. You can assign a cheaper key to lighter work and a stronger one to everything else, and OS3 will not switch providers on its own. Phone calls placed by OS3 need your own OpenAI key and project ID.
Files Stay on the Node, Threads Stay per Channel
rabbit says the local agent does not copy, store, use or sell your files, while conversations and the memory built from them live on rabbit’s servers and task context passes to the model provider you picked under its terms. Messages by iMessage, RCS or SMS now sit in settings next to Telegram and r1, matching what the launch promised. Each channel still keeps its own conversation thread, and rabbit’s what is OS3 page says they share memory but are not one shared chat history.
One month in, OS3 pairs four times the machines it launched with and installs its own agent. That pace is why this launch stuck with our team at TechForward, even with the five-times-faster claim still resting on rabbit’s internal tests.
Sign in at os3.rabbit.tech, then read the pairing guide before you hand it a machine you care about, and track changes on rabbit’s updates page. If local weights are part of your stack, the same discipline from our GPT-OSS coverage applies to the model you serve from your own node.
