The rabbit AI agent is moving beyond its physical constraints, as the startup behind the R1 device has announced a significant pivot toward a software-first ecosystem. On September 22, 2026, Rabbit revealed that its new OS3, described as an “agentic operating system,” will no longer be tethered exclusively to its proprietary hardware. Instead, the system will be available to operate locally across Windows, Mac, and Linux devices, effectively turning existing computers into hosts for Rabbit’s autonomous AI capabilities.
Key Takeaways
- Hardware Independence: Rabbit’s new OS3 agent can run on standard desktop operating systems including Windows, Mac, and Linux.
- Device Flexibility: Users can link up to five different devices to a single Rabbit account.
- Expanded Access: The agent is accessible via desktop sites, messaging platforms like Telegram or iMessage, and the existing R1 device.
- Hardware Pivot: While manufacturing of the R1 has ceased, Rabbit is developing a new “cyberdeck” hardware slated for release within months.
- Privacy Stance: Rabbit claims it will not sell or use user data for training, though it does retain chat histories and “memories” on its servers.
- Model Flexibility: Rabbit states that users can add their preferred AI models to their account. This prevents the company from being locked into a single provider and allows users to choose between speed, intelligence, or cost-effectiveness.
- Local Context: By operating locally on Windows, Mac, or Linux, the agent can see the files you are working on and the apps you have open, providing a level of context that a purely cloud-based chatbot cannot match.
- Multi-Device Orchestration: The ability to link five devices means the agent can potentially coordinate tasks across a laptop, a tablet, and a smartphone simultaneously.
- The Launch of the Cyberdeck: The specifications, price point, and actual capabilities of this new hardware will reveal whether Rabbit is truly committed to a niche hardware strategy or if it is still searching for its identity.
- OS3 Software Rollout: The actual performance of the agentic OS on Windows, Mac, and Linux will determine if Rabbit can transition from a hardware startup to a legitimate software player. Watch for user reports regarding the agent’s ability to handle multi-step tasks and its impact on system performance.
- www.theverge.com
What Happened: The Shift to OS3
In a move that signals a major strategic realignment, Rabbit is transitioning from a hardware-centric company to a software-driven AI provider. As reported by Wired and The Verge, the startup is rolling out OS3, an “agentic operating system” designed to function as a layer of intelligence across a user’s existing digital environment.
While the OS3 system runs in the cloud to process complex logic, it is designed to operate locally on a user’s machine. This hybrid approach allows the agent to interact with files, applications, and hardware directly on Windows, Mac, and Linux platforms. This announcement comes at a critical juncture for the company, following the mixed reception of its initial hardware offering, the R1, which some critics and users described as “underwhelming.”

According to the company, the OS3 system is highly adaptive. It is engineered to automatically determine which specific devices, files, and applications are required to complete a user-assigned task. This level of autonomy is what defines the “agentic” nature of the operating system, moving beyond simple conversational AI toward a system that can execute workflows.
Why It Matters: From Gadget to Ecosystem
This transition is significant because it addresses one of the primary criticisms of the R1: its limitation as a single-purpose, standalone device. By decoupling the Rabbit AI agent from its physical hardware, the company is attempting to capture a much larger market share—those who already own powerful computers and mobile devices but want an intelligent layer to manage them.
For the broader AI industry, Rabbit’s move reflects a growing trend toward “agentic workflows.” Unlike traditional LLMs (Large Language Models) that simply provide text-based answers, an agentic OS is designed to act. If a user asks to “organize my travel itinerary,” an agentic system doesn’t just list flights; it accesses the user’s email, checks their calendar, and interacts with booking applications to complete the task.
By allowing users to access OS3 through familiar interfaces like Telegram or iMessage, Rabbit is lowering the barrier to entry. Users do not need to learn a new way to communicate; they simply interact with the AI through the tools they already use every day.
The Evolution of Rabbit: Hardware and Software
To understand the significance of OS3, one must look at the trajectory of the company. Rabbit entered the market with the R1, a handheld device intended to be a dedicated AI companion. However, the hardware-first approach faced challenges in a market dominated by smartphone giants like Apple and Google, who are rapidly integrating similar AI features into their existing ecosystems.
Jesse Lyu, the founder of Rabbit, told Wired that the startup has officially stopped manufacturing R1 devices. This is not an exit from the hardware space, but rather a refinement of it. Instead of a general-purpose AI gadget, Rabbit is focusing its hardware efforts on a new concept: a “cyberdeck.”
The Cyberdeck and Vibe-Coding
Rabbit is reportedly developing a new piece of hardware centered around “vibe-coding,” a term that suggests a more intuitive, perhaps even more experimental, way of interacting with and programming software through AI. This cyberdeck is expected to launch within months and will serve as a dedicated vessel for OS3.
While the details of the cyberdeck remain sparse, it represents a shift toward specialized hardware that complements the agentic software, rather than attempting to replace the smartphone entirely. This “hardware-as-a-specialized-tool” model may be more sustainable than the “hardware-as-a-replacement” model that the R1 attempted to pioneer.
| Feature | Rabbit R1 (Legacy) | OS3 (New Agentic OS) | Cyberdeck (Upcoming) |
|---|---|---|---|
| Primary Platform | Standalone Hardware | Windows, Mac, Linux, Mobile | Specialized Hardware |
| Access Method | Physical Device | Desktop, Messaging Apps, R1 | Hardware-native |
| Core Function | AI Companion | Agentic Task Execution | Vibe-coding / OS3 Native |
| Device Limit | 1 Device | Up to 5 Devices | TBD |
Understanding the Rabbit AI Agent Architecture
Technically, OS3 operates on a hybrid model. The “brain” of the operation resides in the cloud, allowing it to leverage massive computational power and the latest AI models. However, the “hands” of the agent operate locally on the user’s device.
This distinction is crucial for several reasons:

Privacy and the “Memory” Problem
As AI agents gain the ability to access local files and monitor application usage to perform tasks, privacy concerns naturally intensify. Rabbit has proactively addressed these concerns, though some nuances remain.
According to Rabbit, the OS3 system is designed with a strict policy regarding data usage: the company claims it will not store, copy, use, or sell user data for the purpose of training its own models. This is a significant selling point for enterprise users and privacy-conscious individuals.
However, there is a caveat. Rabbit noted that it will continue to keep user chats and the “memories” stemming from those chats on its servers. In the context of an AI agent, “memories” are essential; for an agent to be useful, it must remember that you prefer certain types of files, that you have a specific meeting every Tuesday, or that you prefer a certain tone in your emails.
Furthermore, because OS3 allows users to integrate third-party AI models, the data privacy landscape becomes fragmented. Rabbit has stated that any other AI providers linked to OS3 will handle user data under their own respective privacy rules. This means a user’s data might be subject to the policies of OpenAI, Google, or Anthropic, depending on which model they have selected for their OS3 agent.
What It Means for You
The implications of the Rabbit AI agent rollout vary depending on who you are in the digital ecosystem:
For the General Consumer
If you are someone who feels overwhelmed by the sheer number of apps and digital tasks on your devices, OS3 could act as a digital concierge. Instead of manually switching between a browser, a calendar, and a spreadsheet, you may soon be able to simply command the agent to “prepare the quarterly report based on my recent emails and save it to my desktop.”
For Tech Enthusiasts and Developers
The introduction of “vibe-coding” and the upcoming cyberdeck suggests a new playground for those interested in the intersection of human intent and machine execution. The ability to run an agentic OS across Linux and Mac makes it a highly flexible tool for developers looking to automate complex local workflows.
For Privacy Advocates
You should expect a more complex relationship with your data. While Rabbit promises not to sell your data, the necessity of “memories” means that a significant amount of your behavioral context will live on Rabbit’s servers. Users will need to weigh the convenience of a highly personalized agent against the reality of centralized data storage.
Counterpoints and Open Questions
Despite the ambitious vision, several questions remain unanswered.
First, there is the issue of security. An agentic operating system that has the authority to “determine the devices, files, and apps it needs to complete a task” is, by definition, a system with deep access. If an agent is compromised, or if it misinterprets a command, the potential for accidental file deletion or unauthorized data exposure is significant. Rabbit has not yet detailed the specific sandboxing or permission-based protocols that will protect local files from the agent’s autonomous actions.
Second, the reliability of execution remains an unproven frontier. While conversational AI is relatively mature, “action-oriented” AI frequently struggles with the nuances of complex software interfaces. Can OS3 reliably navigate a legacy Windows application as effectively as it navigates a modern web browser? The success of OS3 will depend heavily on its ability to handle the “edge cases” of human error and software inconsistency.
Finally, the competitive landscape is daunting. With Microsoft Copilot deeply integrated into Windows and Apple Intelligence coming to macOS and iOS, Rabbit is entering a fight against companies that own the very operating systems Rabbit is trying to layer itself upon.
What Happens Next
In the coming months, the tech industry will be watching for two primary catalysts:

Frequently Asked Questions
Do I need a Rabbit R1 device to use the new AI agent?
No. One of the most significant aspects of the new announcement is that the Rabbit AI agent (via OS3) can run on your existing Windows, Mac, or Linux computers. You can also access it through messaging apps like Telegram or iMessage, making the physical R1 device optional rather than mandatory.
How many devices can I connect to one Rabbit account?
Rabbit has stated that users can add up to five different devices to a single account. This allows for a seamless experience across a desktop, a laptop, and various mobile devices.
Is my data safe with Rabbit’s new OS3?
Rabbit claims that it will not store, copy, use, or sell your data for its own purposes. However, it is important to note that the system will keep your chat histories and “memories” (contextual information used to help the AI) on its servers. Additionally, if you use third-party AI models, those providers will follow their own privacy policies.
What is a “cyberdeck” and when is it coming?
While specific details have not been released, Rabbit’s founder Jesse Lyu has indicated that the company is focusing on a new “vibe-coding” hardware called a cyberdeck. This device is expected to launch within a few months and will be designed to run the OS3 agentic operating system.
Rabbit’s pivot from the R1 hardware to the OS3 software ecosystem represents a high-stakes attempt to redefine the role of AI in our daily computing lives
References
Featured image: Photo by RDNE Stock project on Pexels