Rabbit, the company best known for the R1 handheld, is releasing a standalone AI agent that does not require its own hardware to operate. The Verge reports that the new product, which Rabbit calls OS3, is described by the company as an "agentic operating system" that runs in the cloud but carries out work locally across Windows, Mac and Linux devices.
The shift matters because it decouples the software from the device that defined Rabbit's public reputation. The R1 was widely received as underwhelming, and the new agent is positioned as something a user can adopt without buying anything from Rabbit at all. That is a meaningful change in distribution model for a hardware-first startup.
According to the report, a single account can hold up to five devices, and users can attach their preferred AI models to that account. Rabbit says OS3 will work out which devices, files, apps and models it needs in order to finish a given task, rather than requiring the user to specify the full chain manually.
Access is not limited to the desktop client. The report states that OS3 can also be reached through a dedicated desktop site, through a paired messaging app such as Telegram or iMessage, and on the R1 itself. That gives the agent several entry points, including ones that live inside tools many freelancers already keep open all day.
On hardware, Rabbit founder Jesse Lyu told Wired that the company has stopped manufacturing R1 devices. The report adds that Rabbit is not abandoning hardware entirely: it is focusing instead on a new vibe-coding "cyberdeck" that will run OS3 and is said to launch within "months." No firmer date is given in the supplied material.
The privacy description is narrower than a blanket no-data claim. Rabbit says OS3 will not store, copy, use or sell your data, but the report notes the company will still keep your chats and the memories derived from them on its servers. Other AI providers connected to OS3 handle data under their own privacy rules, which means the effective policy depends on which models a user links.
For freelancers, designers and developers, the practical question is whether an agent that spans several machines and apps reduces context-switching overhead or simply adds another surface to manage. The supported specifics are limited: five devices per account, local operation across three desktop platforms, model choice, and multiple access channels. Pricing, platform-by-platform capability differences, and the exact division of labour between cloud and local execution are not documented in the evidence.
The claim that OS3 "operates locally" while also running in the cloud is the kind of architecture description that usually hides a split: some orchestration or model inference happens remotely, while file and app actions happen on the user's machine. The supplied text does not resolve that split, so anyone evaluating it for client work should treat the local-processing framing as a vendor description rather than a verified technical guarantee.
There is also an unresolved tension in the data story. Saying the system will not store or use your data, while simultaneously retaining chats and derived memories on Rabbit's servers, describes retention rather than absence of retention. The distinction matters for freelancers handling client material under confidentiality terms, because the relevant question is what persists, for how long, and who can reach it.
The cyberdeck is the least defined part of the announcement. It is described as a vibe-coding device that will run OS3, with a launch window of "months" attributed to Wired's reporting. No specifications, price, or confirmed ship date appear in the evidence, so it functions as a stated intention rather than a scheduled product.
The R1's discontinuation is the clearest concrete fact in the report. Manufacturing has stopped, per the founder's statement to Wired, even as the device remains one of the ways to reach OS3. Existing owners are therefore not cut off from the new software, but the hardware line that created the company's audience is no longer being produced.
Read together, the two moves point in one direction: Rabbit is trying to convert a hardware story into a software story, using an agent that runs on machines people already own. That is a lower-friction path to adoption, and it also removes the R1's hardware limitations from the critical conversation, at least for users who never buy the cyberdeck.
The competitive context is not established by the evidence, and it would be a mistake to frame this as a first. What the report supports is narrower: a company with a hardware background is shipping a cross-platform agent with model choice and messaging-app access, and has stopped making its original device.
For working developers and designers, the sensible posture is to test OS3 against a bounded, low-risk task before trusting it with client files. The evidence supports the existence of the product and its stated access methods; it does not support claims about reliability, accuracy, or how well the automatic device-and-app selection performs in practice.
What remains unknown is substantial: pricing, general availability dates, which AI models are supported, how the five-device limit is enforced, what the retention period for chats and memories actually is, and whether the cyberdeck arrives on the timeline described. Until those are documented, OS3 is best treated as an announced system with a clear access model and an unclear operational one.