This Open-Supply Telephone AI Agent Sees, Hears and Acts—All With out Touching the Cloud

Your cellphone already has a digital camera, a microphone, and a display screen. It may well see what you are in actual life and what’s taking place by itself show. And now, the AI group from Chinese language smartphone producer Oppo has discovered that each one that {hardware} that sits there, principally underused, is precisely what you should construct a genuinely helpful cell AI agent.

That undertaking is X-OmniClaw, revealed by the Multi-X Workforce. It is an open-source AI agent framework for Android that turns your cellphone right into a hands-free, context-aware assistant able to working actual duties throughout actual apps, with out routing every part by means of a cloud copy of your system.

Most cell AI techniques do not really run in your cellphone. They run on cloud servers that host digital copies of Android, letting an AI faucet and scroll by means of apps remotely. The outcome: no entry to your actual digital camera, your precise pictures, or your native information—only a stranger utilizing a duplicate of your cellphone.

X-OmniClaw takes the alternative method. Per the technical reportit introduces “an edge-native structure that executes straight on the person’s bodily system, thereby eliminating the hole between simulated environments and real-world interplay contexts.”

The report makes use of a automobile analogy: The smartphone is “the automobile,” X-OmniClaw is “the interior engine for management and notion,” and the cloud-based language mannequin is simply known as in as “the gas” when heavy reasoning is required. All the things else stays native.

How the Oppo AI cellphone agent works

X-OmniClaw’s general structure relies on three pillars: Omni Notion, Omni Motion, and Omni Reminiscence that work as one steady loop, with cloud LLMs known as in just for heavy reasoning, in line with Oppo.

Omni Notion covers every part the cellphone can sense. It combines digital camera feeds, display screen content material, and voice enter right into a single pipeline. A vision-language mannequin interprets the scene earlier than the agent does anything. So in case you level your digital camera at a bottle and ask, “how a lot does this value?”, the agent first figures out what you are , then opens the related buying app and begins looking out. No guessing required.

Omni Reminiscence is what separates X-OmniClaw from a one-shot chatbot. The agent maintains context throughout duties, app switches, and classes. It additionally builds a long-term semantic reminiscence out of your picture gallery, turning uncooked photographs into structured notes about objects, scenes, and occasions. The report states “runtime continuity is what lets X-OmniClaw function as an ongoing system agent somewhat than a one-shot response system.”

Omni Motion handles execution. It combines XML interface information with an on-device visible mannequin and OCR—a character-recognition layer to determine precisely what to faucet, even on ad-heavy screens the place construction alone is not sufficient. It additionally consists of conduct cloning: report your self navigating to a buried app web page as soon as, and the agent can replay that route immediately utilizing an Android deeplink shortcut subsequent time.

What the Oppo AI agent can really do

Oppo shared some issues the mannequin can do. For instance, the agent identifies a bodily product through digital camera, opens Taobao, scrolls outcomes, and returns a worth abstract—no typing required.

Oppo additionally demoed a floating on-screen companion that helps a person work by means of math workouts step-by-step: autonomously studying the display screen, processing every query, and advancing when executed.

It additionally supplied one other instance by which a person asks the agent to assemble a spotlight video from parrot-themed pictures. The system scans the gallery, finds matching pictures utilizing its semantic reminiscence, opens CapCut’s video editor through deeplink, batch-selects the information, and generates the video. What used to take “a couple of minutes or longer” turns into a handful of automated steps.

2026: The 12 months of agentic AI

AI brokers have change into one of the vital mentioned classes in tech. OpenClaw—the open-source agent framework that reached over 373,000 GitHub stars and was ultimately backed by OpenAI—launched the present wave by displaying what persistent, locally-run brokers might do on PCs. Hermes Agent by Nous Analysis took issues additional with a self-improving learning loop that compounds capabilities over time.

Each run totally on desktop {hardware}. X-OmniClaw extends the identical structure to the system you really carry in every single place. The group constructed on the open-source HermesApp codebase, and the paper explicitly credit OpenClaw’s structured skill model as foundational inspiration, then tailored it for the multimodal, always-on nature of a smartphone.

The code is on GitHub now. Oppo says it’ll launch all property and hold updating the undertaking because the system evolves.