Oppo’s new open-source phone agent turns your smartphone into a privacy-first, always-on assistant — no cloud tether required.
The Multi-X Team at Chinese phone maker Oppo has published X-OmniClaw, an open-source AI agent framework for Android that runs directly on your physical device. Rather than routing interactions through cloud-hosted “virtual phones” (which can’t access your real camera, files or local apps), X-OmniClaw executes perception, memory and actions at the edge, calling cloud language models only when heavy reasoning is needed.
Edge-native architecture, not simulation
Oppo describes X-OmniClaw as closing “the gap between simulated environments and real-world interaction contexts.” Using a vehicle analogy from the technical report: the smartphone is the vehicle, X-OmniClaw is the internal engine for control and perception, and cloud LLMs are simply fuel you call on occasionally — most sensing, planning and control stays local.
Three pillars: Omni Perception, Omni Memory, Omni Action
- Omni Perception merges camera feeds, screen content and voice input into a single pipeline. A vision-language model interprets the scene before any step is taken, so the agent can reliably identify a real-world object (say, a bottle) and then open the correct shopping app to search for price information.
- Omni Memory gives the agent continuity across tasks, app switches and sessions. It builds a long-term semantic memory from your gallery and other data, turning raw photos into structured notes about objects, scenes and events so the assistant remembers context rather than answering as a one-off.
- Omni Action combines Android XML interface metadata with on-device visual models and OCR to decide exactly what to tap, even on ad-cluttered screens. It also supports behavior cloning: record a navigation once and the agent can replay it later via Android deeplinks.
Real-world demos, zero typing
Oppo’s demos illustrate everyday automations made possible by on-device perception and control. Point the camera at a product, and the agent can identify it, open Taobao, scroll results and return a price summary without you typing. A floating on-screen assistant can autonomously read and solve step-by-step math problems by reading the display and advancing when a question is complete. For media tasks, the agent can scan a photo gallery for “parrot” images using its semantic memory, open a video editor by deeplink, batch-select files and assemble a highlight reel — cutting minutes of manual work into an automated sequence.
Builds on recent open-source agent work
X-OmniClaw stands on the shoulders of the recent wave of locally run agents. OpenClaw — the now-popular open-source agent framework that hit 373,000 GitHub stars and later received OpenAI backing — helped launch the trend on PCs. Hermes Agent introduced self-improving loops and capability compounding. Oppo adapted those desktop ideas for the always-carried, multimodal smartphone context, building on the HermesApp codebase and crediting OpenClaw’s structured skill model as core inspiration.
Open-source, on GitHub, and evolving
The code for X-OmniClaw is available on GitHub now. Oppo says it will publish all assets and continue updating the project as it evolves.
Why this matters for privacy and decentralization-minded readers
For crypto and privacy communities, X-OmniClaw’s emphasis on local execution is notable: it reduces dependence on cloud copies of your device and the data flows that come with them. Running perception and control on-device better preserves user data sovereignty and aligns with the broader decentralization impulse to keep sensitive processing in the hands of the owner. For developers, an open-source, phone-native agent that can operate across apps and media opens new possibilities for secure, private automation and richer on-device UX.
Bottom line: X-OmniClaw brings persistent, multimodal agents to the phone you actually carry, prioritizing local sensing and control and reserving the cloud for heavy reasoning — an open, privacy-oriented step toward more capable personal agents.
Read more AI-generated news on: undefined/news
