Rabbit OS3 moves the agent off the R1 and onto Windows PCs
In its launch announcement, Rabbit says it today announced the general release of OS3, an agentic operating system. It is built around one chat interface. A user states an outcome, and the system works out the rest by reasoning with connected AI models and whatever else the task calls for: the web, services already in use, or connected devices.
The Verge covered the launch after WIRED's earlier report. It called the R1 "underwhelming" and described OS3 as a standalone AI agent that you don't need its hardware to use. That is the main change. Rabbit's earlier software revolved around the R1, a small device that runs rabbitOS, an operating system based on the Android Open Source Project. Under OS3, as the Dubai tech site tbreak puts it, the R1 is now one endpoint among several.
The R1 is still supported. Rabbit's list of controllable devices covers Windows, Mac, or Linux local PCs, cloud VMs, dedicated AI machines, and the r1 today, with more device types to follow. It lists these ways to send commands: the web portal on desktop or mobile, a paired Telegram, iMessage/RCS/SMS, or with rabbit's own r1. The company says more input methods are coming but hasn't named them.
Rabbit has been building toward this for a while. It points to twelve consecutive months of upgrades delivered for rabbit devices, including rabbitOS 2, terminal mode, and a move to bring-your-own-key support for DLAM. rabbitOS 2 came out in September 2025 and introduced a redesigned card-based interface, gesture navigation, and a "creations" feature for generating small software tools and experiences on the device.
How OS3 splits work between the cloud and the local Rabbit agent
The Verge summed up OS3 as software that "runs in the cloud but operates locally." Rabbit's own description is more precise. OS3 runs in the cloud, but connects to and operates devices through the rabbit agent, which installs with a single command. The planning, the memory and the conversation live on Rabbit's servers. Each connected PC runs a small agent that carries out actions on that machine.
Rabbit says one account can be connected to up to five devices, and OS3 determines where each task runs (on a single device, or across several). It also says the system moves work between devices mid-task and pulls in any files, apps and skills it needs. In practice, a task that starts on one machine can finish on another. Crypto Briefing describes the pattern this way: a task that starts on your Mac can be continued on a Windows machine or handed off to a cloud instance. The local agents running on each device handle direct control, while the central chat interface serves as the single point of interaction.
Actual control of the desktop comes from Rabbit's action models. Using rabbit's DLAM technology in conjunction with a local agent, OS3 can operate desktop software, local files, and web pages directly on a user's behalf, in addition to writing, debugging, and running software workflows autonomously. Crypto Briefing says this means the system interacts with software the way a human would, by navigating the interface itself, so apps don't need purpose-built integrations. That approach dates back to the R1, which was sold on a "large action model" that the company described as software able to operate websites and services on behalf of users.
The interface drops separate chat threads. The company announced the general release on 22 September, describing OS3 as an agentic operating system built around one continuous chat rather than separate sessions. Rabbit says the system keeps track of earlier outputs and remembers what you've told it, which matters for the privacy question covered below.
Rabbit hasn't published the practical details a Windows administrator would want. The announcement gives no supported Windows editions, minimum hardware, agent version or step-by-step install procedure beyond calling it a single command. Setup starts in a browser: users can get started by visiting os3.rabbit.tech via desktop to log in or create a rabbit account, and finish onboarding in a single conversation.
OS3's bring-your-own-key models and pasted skills keep it open
Rabbit's main pitch against the big platform vendors is openness. Users can bring API keys from major AI labs, cloud routers or locally hosted models, then swap models without changing their context, memory or installed skills. In Rabbit's model, the AI provider is a replaceable part and Rabbit supplies the setup around it: identity, memory, device control and skills.
Adding skills is meant to be simple. Rabbit says any published skill works: just copy the skill's URL, paste it into the chat, and OS3 will set it up without requiring command line prompts or additional configuration files. The company presents this as convenience. On a Windows PC it also means new capabilities can be added to an agent with desktop access in one paste, with no installer or configuration file you might stop to review first. Rabbit's materials don't describe any vetting or sandboxing for these third-party skills.
Crypto Briefing frames the openness as a strategic bet. It says Rabbit is betting that none of the major platform incumbents will build a truly platform-agnostic agent. Microsoft has no incentive to make its AI work seamlessly on macOS. That is the outlet's analysis, not Rabbit's. It is a plausible reading of why a 15-person startup would launch on three desktop operating systems at once. tbreak places OS3 in a busier category that already includes products such as OpenClaw's shared browser sessions and ChatGPT Agent. Rabbit's difference is the attempt to keep one agent identity across local computers, cloud machines and messaging apps.
All the capability claims so far are Rabbit's own. No outlet has published independent benchmarks, task success rates or a list of supported models. tbreak also noted that the announcement does not list a subscription price. If you use your own API keys, you will at least pay whatever your model provider charges, on top of any price Rabbit may set for OS3 itself.
Where OS3's "files stay local" promise ends
The Verge's summary is that Rabbit says OS3 won't store, copy, use or sell your data but does keep your chats and the memories built from them. Rabbit's full statement draws a finer line, and Windows users should understand it before installing the agent.
Rabbit's first claim is about the local agent: The rabbit agent runs locally on your machine and does not copy, store, use, or sell your data. The second is about conversations: your conversations with OS3 and the memory built from them are stored on Rabbit's servers, while the files on your computer are not. The third, in the announcement's section on model providers, is the one that matters most. OS3 does its work by sending relevant content to the model provider you choose, which then handles it under its own terms. Before that handoff, the data is processed through Rabbit's servers along with any prompts needed for the task. Rabbit says it doesn't keep a copy of that data and the files themselves stay on your computer.
Put together, the files stay on your disk, but content from those files can leave the machine. When a task needs a document, the relevant portions go to Rabbit's cloud and then to OpenAI, Anthropic, a routing service or whoever holds the API key you supplied. Anything you say in the chat also becomes stored memory on Rabbit's servers. If you point OS3 at a locally hosted model, which Rabbit supports, you control the provider side of that chain. Rabbit's announcement still describes content being processed through its own servers first.
On safety, Rabbit says OS3 only acts when you tell it to, sensitive actions require your confirmation, the local operating system handles system-level permission requests, and you can revoke permissions at any time. The announcement doesn't say which actions count as "sensitive," how long chat memory is kept, what encryption is used, or whether there are admin controls, audit logs or compliance certifications. For a personal PC, those gaps mean you're relying on Rabbit's word. For a managed Windows fleet, they mean OS3 has nothing yet that an IT department could formally evaluate.
The R1 winds down as Rabbit pivots to the Cyberdeck
OS3 also marks a hardware change at Rabbit. According to The Verge, though Rabbit founder Jesse Lyu tells Wired that the startup has stopped manufacturing R1 devices, it isn't giving up on AI hardware entirely. It's instead focusing on a new vibe-coding "cyberdeck" that will run OS3 when it launches within "months," Wired reports. Rabbit's own announcement confirms the Cyberdeck plan. It describes the device as the default operating system on the company's upcoming Cyberdeck and calls it a dedicated "vibe coding machine." The announcement doesn't mention ending R1 production or give a Cyberdeck release date. Both of those details come from Lyu's interview with WIRED.
tbreak reports more figures from WIRED. WIRED's report says Rabbit has delivered more than 100,000 R1 units, despite the device receiving a 3/10 review from the outlet at launch. Rabbit founder Jesse Lyu told WIRED the company has raised around $60 million and employs 15 people. Those numbers help put the pivot in context. The Windows, Mac and Linux PCs people already own give Rabbit far more potential users than about 100,000 handhelds.
Lyu presents OS3 as the goal Rabbit was working toward all along. "We've spent two and a half years building toward a system you don't operate, but instead just tell it the outcome."
What this means for Windows users weighing OS3
If you're curious about OS3, try it on a machine you can afford to expose, not your main work PC. Organizations should keep it off managed endpoints until Rabbit publishes retention, administration and security details. An agent that can operate desktop software and read local files is effectively a remote-control tool driven by a cloud service and a third-party AI model. Its installer also runs from a single command that Rabbit hasn't documented in detail.
R1 owners have less to worry about. The device is still an OS3 input and a controllable endpoint, so it keeps working, but Lyu's statement to WIRED means no new R1 hardware is coming.
- OS3 became generally available on September 22, 2026. Setup starts at Rabbit's OS3 web portal on a desktop browser, and no R1 is required.
- Each PC you want OS3 to control has to run the local Rabbit agent, and one Rabbit account can connect up to five devices, including Windows, Mac and Linux PCs, cloud VMs and the R1.
- Content from your files is processed on Rabbit's servers and sent to your chosen model provider, and your chats and derived memory are stored by Rabbit, even though the files themselves stay on your PC.
- Bring-your-own-key support lets you choose frontier, routed or locally hosted models, so check your provider's data terms and API costs as well as Rabbit's.
- Skills installed by pasting a URL add capabilities to an agent with desktop access, so only add skills from sources you trust.
- Rabbit hasn't published a price, supported Windows editions, hardware requirements, retention periods or enterprise controls, so budget and compliance decisions will have to wait.
OS3 changes what Rabbit is. The company started with a handheld built around an action model and now offers that model as software for machines people already own, with Windows among the first platforms. Its capabilities and its privacy promises are still Rabbit's own claims, and its data flow runs through Rabbit's cloud and at least one outside AI provider. The next concrete step is the Cyberdeck, which WIRED says is due within months with OS3 as its default operating system. By then, Rabbit should have published pricing and administration details for anyone considering OS3 on more than a test machine.