London – September 09, 2026 -- TrustKernel has moved its PlugClaw device from pre-order to general availability worldwide, pricing the thumb-sized private AI computer at 128 EUR ($149) and adding free confidential model inference on newly deployed servers.
PlugClaw packs a full computer into a 50 x 19 x 8mm USB-C stick weighing 15 grams
The device runs a MediaTek Helio G80 processor with up to 6GB of RAM and 128GB of encrypted storage. It ships with PlugOS, a privacy-hardened Android build, alongside Ubuntu as an agent runtime environment, turning any connected smartphone, tablet or computer into an AI workstation.
The hardware isolates AI agents from host device data by design
When plugged into a phone, tablet or PC, PlugClaw uses the host purely as a screen and input device. The agent, its apps, credentials and working data execute entirely inside PlugClaw's own compute environment without requiring administrator access to the host, and the agent cannot reach the owner's photos, documents or accounts. TrustKernel points to prior incidents where apps exceeded granted permissions and exposed user data as the rationale for the isolated architecture. PlugClaw natively runs OpenClaw with an added GUI agent layer and Android-use capability, letting it open apps, respond to messages, organize files and execute multi-step tasks confined to the device.
The stick supports seven operating systems from a single hardware unit
PlugClaw works across Android, iPhone, iPad, Windows, Mac, Linux and HarmonyOS, presenting the same apps, storage and workflows regardless of which device it is plugged into.
Frontier-model requests route through trusted execution environments that block data access
Tasks requiring frontier AI models are processed in cloud-based hardware running Trusted Execution Environments, structured so that neither the cloud provider nor TrustKernel can view the data. In pre-launch testing, users chose the confidential model option over a cheaper, non-confidential alternative significantly more often. TrustKernel has now provisioned dedicated servers offering this confidential inference free of charge.
Buyers get a one-time hardware purchase with no subscription and a $15 usage credit
Users can bring their own API keys or use pre-installed models, with a $15 credit included at purchase. Test units are available on request.