A new wave of Chinese products and chip designs is racing to make devices more than passive endpoints for cloud models. Sparked by OpenClaw — an open, self‑hosted agent framework nicknamed “Longxia” (literally "lobster") — manufacturers from smartphone giants to robot makers are adding on‑device agents that can act autonomously across apps and peripherals rather than merely reply to queries.
OpenClaw’s selling point is simple but transformational: run a contextual, persistent AI agent on local hardware with full, user‑granted access to system services. That changes the interaction model from the single‑shot, cloud‑driven “ask and respond” of familiar voice assistants to multi‑step workflows that can schedule meetings, move files, control sensors, and trigger actuators without round trips to remote servers.
The market reaction in China has been immediate. Vendors large and small are branding “Claw” features — Xiaomi’s miclaw, Huawei’s XiaoyiClaw, OPPO’s XiaobuClaw and Baidu’s DuClaw — while chipmakers such as Haiguang, Loongson and Rockchip tout local support. Mac mini 16GB versions have reportedly sold out in response to demand for local agent hosting, and cloud providers including Alibaba, Tencent and Baidu offer one‑click OpenClaw deployment to ease developer and consumer adoption.
Beyond phones, the technology is being shoehorned into wearables, eyewear and robots. Smart‑glasses makers are piping camera and microphone streams into local agents to enable first‑person perception; robot companies are using OpenClaw to give machines closed loops of perception, planning and action — a robot dog that can “look, move, observe and report” on command is one early example. These use cases underline how local agents bridge large models’ reasoning power with hardware-specific abilities to effect change in the physical world.
That fusion rewrites the rules of competition for hardware. Historically, vendors competed on sensors, chassis and spec sheets; now the axes include on‑device models, local compute efficiency, fine‑grained hardware permissions and the APIs that let agents reach into system services. Several Chinese AI entrepreneurs and industry figures have dubbed the agent layer — when running on the edge — as the next operating system, because it mediates users’ intents across disparate apps and devices.
The promise is accompanied by clear caveats. Running powerful agents locally expands the threat surface for data leakage, privilege escalation and supply chain tampering, and it raises questions about update regimes for on‑device models and workflows. Companies and security experts have urged caution: early adopters are recommended to test in controlled environments and back up data before upgrading primary devices.
Globally, OpenClaw’s rise illustrates two broader trends. First, decentralisation: major functionality is migrating from cloud APIs back onto endpoints as models and accelerators become small and efficient enough for phones and edge chips. Second, platformization: whoever owns the agent runtime, the model zoo and the hardware bindings gains leverage over a new ecosystem of apps and services. For incumbents and regulators alike, that creates both opportunity and headache — from fresher consumer experiences and new hardware replacement cycles to regulatory scrutiny over safety, privacy and cross‑border model governance.
The technology remains nascent, but its direction is clear. If OpenClaw and its imitators deliver reliable, secure local agents, the next two years could see a rapid consumer upgrade cycle and a reshaping of supply chains and software stacks. If they fail on safety or interoperability, vendors who rushed in may face backlash and slow adoption. Either way, the industry is in motion: "agentified" devices are no longer a fringe experiment but a defining vector for the next generation of AI hardware.