Mylobster.ai plans kernel-level AI assistant platform

Mylobster.ai has laid out a plan for personal AI assistants that run with operating system-level access rather than as conventional applications, arguing that the next major platform shift in consumer and workplace computing will sit close to the kernel.

In a blog post, founder Dendi Suhubdy said today's assistants and agent products are constrained by "userspace" access, where software operates through published APIs, accessibility tools, and integration tokens. He argued that working at a deeper layer would change what assistants can observe and do across devices.

He pointed to the recent popularity of OpenClaw, an open-source personal agent that has spread rapidly among developers and online communities. Suhubdy described it as evidence of demand for automation that can perform actions on a user's machine, even if it runs as a Node.js daemon in userspace.

Kernel focus

The thesis rests on the kernel's position as an operating system's most privileged layer, with visibility into system calls, file operations, process activity, and network traffic. An AI layer placed there would gain system awareness that typical applications lack.

Suhubdy said many assistants face structural limits. Observation is incomplete when an agent only sees what a user pastes in or what integrations expose. Action is indirect when it must pass through multiple API layers. Persistence can also be fragile for always-on agents that compete for resources and may be killed under system pressure.

The post also tied the idea to market size. It cited a personal assistant market that includes Siri, Alexa, and Google Assistant, valued at roughly $15 billion and projected to reach $90 billion by 2030. It argued that system-level assistants could expand into adjacent categories such as IT support, productivity software, monitoring and observability, cross-device synchronisation, and endpoint security-an opportunity Suhubdy characterised as $200 billion-plus.

Product outline

Mylobster.ai described its intended product as "a vertically integrated platform that embeds AI at the operating system level across every major platform - iOS, Android, Windows, macOS, Linux, and web - connected through an encrypted peer-to-peer mesh." The post positioned it as distinct from a chatbot interface.

On desktop systems, the design includes a privileged component. For Linux, it described loadable kernel modules and eBPF programs hooking into filesystem, network, process, and security subsystems. For Windows, it described kernel-mode drivers using KMDF with filesystem and network filtering, plus ETW integration. For macOS, it described system extensions using Apple's Endpoint Security and Network Extension frameworks.

Alongside the OS component, the company outlined an encrypted mesh linking a user's devices via a WireGuard-based peer-to-peer network. It said the approach avoids a central server for raw telemetry and behavioural history, and that only "carefully scoped, distilled prompts" would be sent to large language model providers for inference.

For the agent core, the post referenced a customised distribution of OpenClaw that would take a stream of events from the kernel layer as an additional input source, alongside messages and other interfaces.

Use cases

The post listed examples it said kernel-level visibility and a device mesh would enable, including context-aware prompts during software development, continuity as a user moves from a desktop to a phone, and network troubleshooting based on packet-loss detection and calendar context. It also described monitoring memory behaviour to flag potential leaks and observing patterns such as typing speed and application switching.

It argued a kernel-level approach could create a competitive barrier because kernel development is difficult and mistakes can crash systems. It added that cross-platform support multiplies testing complexity, and that a multi-device mesh could increase switching costs through accumulated context.

Security questions

Suhubdy addressed privacy and security concerns directly, acknowledging that kernel-level visibility may raise objections. He argued that many consumers already grant extensive access to platform vendors through smartphones and cloud services, and positioned Mylobster.ai's architecture as keeping raw data on user devices.

The post described a security model that includes signed kernel components verified at boot, eBPF verification on Linux, and an approach in which kernel components "observe and report" while actions require explicit authorisation. It also referenced third-party security audits with published results and said the company plans to document its security model publicly.

It also outlined a trade-off between local inference and cloud model quality, saying local models lag frontier models and that many users may prefer cloud inference, while claiming the boundary design limits what providers can see.

"The question is not whether AI integrates at the kernel level. The question is who builds it," Suhubdy said.

Mylobster.ai is recruiting kernel engineers, networking engineers, AI infrastructure engineers, and security researchers. Suhubdy argued that the window for independent entrants sits between viral proof points in agent software and deeper moves by Apple, Microsoft, and Google.