Rabbit OS3 Goes Live Without the R1, Turning Your Own PC Into the Agent
Rabbit OS3 is the company's new agentic operating system, and it ships without the hardware that made Rabbit famous. The Santa Monica startup released the platform to general availability on September 22, 2026, after an invite-only beta, pitching it as software that runs in Rabbit's cloud while doing the actual work on the Windows, Mac and Linux machines a user already owns.
Installation happens through a single command that places a local agent on each computer. Onboarding runs from a desktop browser at os3.rabbit.tech, where a new user logs in or creates a Rabbit account and finishes setup in one conversation. From then on the agent is reachable through a web portal on desktop or mobile, through Telegram, iMessage, RCS or SMS, or through the R1 handheld.
How Rabbit OS3 Splits Cloud From Machine
The design divides labour in two. Rabbit's cloud handles orchestration and memory, while the locally installed agent touches files, applications, terminals and supported screen controls on each paired machine. Rabbit states that OS3 works out on its own which connected computer should perform a given task, so a request typed into Telegram can end up executing on a desktop in another room.
Rabbit lists three core capability groups: web research, file and browser work, and code. The company says the agent can write and run code, then debug it, an approach it describes as a Direct Large Action model. The platform also advertises Universal Skill Compatibility, which lets users plug in preferred AI models rather than accepting one fixed provider.
One account can link up to five computers. Chats and related memories stay on Rabbit's servers, and the company says it does not sell user data.
| OS3 detail | What Rabbit states |
|---|---|
| Availability | General release, September 22, 2026 |
| Supported machines | Windows, Mac, Linux |
| Device limit | Five paired computers per account |
| Access channels | Web portal, Telegram, iMessage, RCS, SMS, R1 |
| Model access | Bring your own key; locally hosted models supported |
| Subscription | None |
Rabbit's label of agentic operating system deserves scrutiny. OS3 does not replace Windows, macOS or Linux and does not boot anything. It is an orchestration layer with a memory store and a set of device agents. The name signals ambition rather than architecture, and the practical test is whether the layer can hold enough context about a user's files and habits to justify routing everyday tasks through it instead of doing them by hand.
The permission model is the part consumers should weigh most carefully. A local agent that can read files, drive a browser, run terminal commands and control screens has broad reach over a personal machine, and Rabbit has published few technical details on how it isolates tasks, what the agent touches by default, or how a user revokes access.
How Rabbit OS3 Differs From the R1 Pitch
The R1 was a purpose-built handheld meant to stand in for a phone. OS3 inverts that: the agent is the product, and the hardware is optional. OS3 reaches the R1 as a software update, but the handheld now is one input and output channel among six rather than the entry point. R1 production has wound down as Rabbit leans on software, and R1 owners route their queries through OS3 after onboarding.
Rabbit has also cleaned house on the older software. Data from the earlier rabbithole carries over to OS3 accounts, while the old rabbithole and the standalone rabbit intern have been retired.
The strategic reading is straightforward. A dedicated AI gadget carries manufacturing cost, a fixed form factor and a narrow addressable market; a cloud agent that rides on hardware people already own carries almost none of that. Rabbit spent roughly 50 software updates building the assistant layer on the R1, and OS3 sells that layer on its own.
Reaching users through Telegram, iMessage, RCS and SMS is a distribution decision as much as a convenience one. It means no app to download and no new habit to learn, and it puts the agent inside threads people already open dozens of times a day. The cost is a thinner interface: rich tasks such as reviewing a codebase or reorganising a drive land better on the web portal than in a text thread.
Pricing, Limits and the Open Questions
Rabbit's bring-your-own-key model is the sharpest commercial difference from subscription rivals. Users supply their own model API credentials, so Rabbit charges nothing monthly and the running cost falls to whichever provider the user picks. That arrangement shifts price risk onto the user: heavy agent use bills against an OpenAI, Anthropic or Google key, and a locally hosted model removes that bill only if the user owns the hardware to run it, which for most consumer laptops is not the realistic path.
Two limits stand out. The five-device cap constrains anyone trying to cover a work laptop, a home desktop, a second office machine and a spare. And because memory and conversation history live on Rabbit's servers, the agent's usefulness depends on a cloud account that a self-hosted setup would avoid.
Availability is uneven as well. Rabbit has not confirmed country-by-country rollout or regional pricing, which leaves buyers outside the United States without a clear answer on access. The company has also said nothing about where non-US account data would be stored.
Who gains most from OS3 today is a narrow group: people who already run several machines and want one place to hand off research, file shuffling and small scripting jobs. Anyone expecting the phone replacement that the R1 originally promised gets a browser tab and a chat thread instead.
Why this matters
For consumers, OS3 turns an existing PC into an agent endpoint without a hardware purchase, which makes the category cheaper to try than any AI gadget shipped so far. The trade-off is that the intelligence is rented: models come from a key you supply, and memory sits on Rabbit's servers. The next signal to watch is whether Rabbit ships its referenced cyberdeck, because that would test whether the company still believes hardware has a place alongside the agent layer.
Related Articles
- Google Debuts Gemini Omni Flash and Managed Agents for Enterprise Automation
- OpenAI Agents API Public Beta Moves Agent Orchestration Onto OpenAI's Servers
- Huawei Cloud AICS Goes Global as Agentic Stack Targets Regulated Enterprises
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.