
Share
Shanghai Droi Technology is baking AI into the operating system itself rather than bolting it onto apps, with a hybrid cloud/on-device architecture now shipping on $160 phones from Coolpad and Philips.
Most "AI phones" right now are just regular phones with a chatbot bolted on. You still open the camera app to edit a photo, still open a separate assistant to summarize a document, still ping-pong between apps to book a flight. The AI is a feature, not the foundation. DroiClaw, a project from Shanghai Droi Technology (branded Zhuge in some markets), is trying something structurally different: push AI down into the OS layer itself and see what happens when apps become optional.
The pitch is an "Agentic OS", a system where you describe what you want in natural language and the operating system figures out how to execute it, rather than you hunting through a grid of icons to find the right app for the job. That's a meaningfully different design target. A traditional OS asks "which app do you want to open?" DroiClaw is trying to get the system to ask "what are you trying to accomplish?" instead.
Here's the architectural distinction that matters for anyone who's shipped software on top of Android or iOS: most AI features today live in userspace. They're apps, or SDKs inside apps, calling out to a model API. DroiClaw's claim is that AI functions are wired into system scheduling, task execution, and resource management directly, so the assistant isn't a thing you launch, it's part of how the kernel-adjacent layers coordinate work.
Practically, that means an agent can:
Sensitive operations still require explicit authorization, which is the right call. Full autonomous execution without a permission gate is how you end up with an agent that books a flight to the wrong city or deletes a file it shouldn't. DroiClaw's model keeps a human in the loop for anything high-stakes while letting routine, low-risk workflows run end-to-end.
On the inference side, DroiClaw uses a hybrid architecture rather than betting everything on either on-device or cloud compute. Under normal conditions, the system leans on cloud-based large language models for the heavy lifting: multimodal interaction (voice, text, image, video, file inputs), AIGC content generation, complex reasoning, and skill execution. That's the sensible default, since cloud models are simply more capable than anything you can currently run efficiently on a sub-$200 device.

But when connectivity drops or cloud credits run dry, the system falls back to an on-device model for basic Q&A. It's not a seamless capability swap, the on-device fallback is explicitly more limited, but it's a reasonable degradation path rather than a hard failure. The platform also supports multiple cloud model backends, including custom and privately deployed options, so device makers and developers aren't locked into a single model vendor.
That multi-model flexibility is worth flagging for anyone who's dealt with vendor lock-in on AI infrastructure. An OS-level AI layer that only works with one provider's model is a liability the moment that provider changes pricing or gets deprecated. Supporting swappable backends is a more defensible long-term architecture, even if it adds complexity to the orchestration layer.
On the trust and safety side, DroiClaw leans on permission management, task scheduling boundaries, and layered data protection to keep agent actions auditable. The stated goal is operational visibility, so users (and presumably developers building skills on top of the platform) can inspect what an agent actually did and why. Combined with local processing for privacy-sensitive scenarios, it's a reasonable attempt at the transparency problem that's going to dog every agentic system as these things get more autonomous. An agent that silently chains together a dozen actions on your behalf is only useful if you can later audit exactly what happened.
The first commercial hardware running DroiClaw comes from Coolpad and Philips, priced around RMB 1,099, roughly $160. That's notable mostly because it's not a flagship play. A lot of "AI-native" positioning in this industry has shipped first on premium hardware, Pixel, Galaxy S-series, that kind of thing, with advanced features gated behind top-tier chipsets. Shanghai Droi Technology is explicitly going the other direction, framing broad access as core to the strategy rather than a later phase.
The company isn't new to this space. Shanghai Droi Technology was founded in 2008, and its earlier platform, FreemeOS, has reportedly been deployed on more than 200 million devices through a partner network of over 1,000 hardware and software companies across Europe, India, Southeast Asia, and Latin America. That track record matters for credibility here: building an OS layer that ships reliably across a fragmented hardware ecosystem is a genuinely hard distribution problem, and FreemeOS's footprint suggests the company has already solved a decent chunk of it.
DroiClaw builds on that distribution base rather than starting from zero. Users can personalize AI avatars, install and create custom "skills," set scheduled tasks, and configure which underlying models handle which requests. Whether any of that translates to a genuinely app-free experience in daily use, or just a smarter overlay with extra branding, is the open question nobody outside Shanghai Droi Technology can answer yet.
An app-free OS is a big structural bet, and the industry has seen ambitious platform rewrites stall before. But the architecture here, system-level AI integration, hybrid inference with graceful fallback, multi-model support, and a push into budget hardware instead of flagships only, is a coherent response to where agentic computing seems to be heading. Worth watching whether third-party developers actually build meaningful "skills" for it, since an agentic OS is only as useful as the breadth of tasks it can coordinate. The 200-million-device FreemeOS legacy gives Shanghai Droi Technology a distribution advantage most AI-OS challengers don't have, but distribution alone won't settle whether users actually want their phone to stop asking "which app" and start asking "what do you want."
Tags
Original Sources
How DroiClaw is building an AI-native operating system for the agentic era | TechCrunch
↗ https://techcrunch.com/sponsor/droiclaw/how-droiclaw-is-building-an-ai-native-operating-system-for-the-agentic-era
About the author
Kai built ML infrastructure at a Bay Area startup before developing an obsession with transformer architectures and inference optimisation that eventually pulled him out of product work entirely. A stint at a compute research lab sharpened his instinct for what actually matters in a model release versus what is marketing. He writes from the inside — from the perspective of someone who has debugged the systems he is describing at three in the morning. He is allergic to hype and instinctively drawn to the unglamorous plumbing questions that everyone else skips over.
More from The Engineer →This Week's Edition
13 November 2026
14 articles
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
Related Articles

Tether Evo's New Research Tackles BCI's Calibration Problem with Cross-Subject Models
Models & Research · 5 min

Why Some Engineers Think the Ocean Floor Is the Future of Data Centers
Environment & Climate · 5 min

Europe's Digital ID Push Tests a New Privacy Tool: Zero-Knowledge Proofs
Policy & Regulation · 5 min
More Stories
© 2026 Cedar & Bloom. All rights reserved.