A notification-driven mobile agent with shared local context and user-confirmed actions.
Paper · Poster · Demo video · Download APK · Quick start
MobileClaw is an Android research prototype developed at the Intelligent Systems Lab (PittISL), University of Pittsburgh. Notifications can initiate contextual assistance, and foreground chat uses the same phone-local conversations, memories and tool results. The phone coordinates the agent loop and executes tools through Android interfaces.
MobileClaw uses a cloud model service by default and can connect to any OpenAI-compatible model server, including locally hosted models and servers running on the phone over loopback. It does not bundle an LLM or an inference engine.
MobileClaw-publish.mp4
Download the demo video (MP4, 1920 × 1080, 2 min 59 sec, English narration and subtitles). See the demo setup guide for reproducible examples.
Join us for an in-person demonstration of MobileClaw at ACM MobiCom 2026.
- When: Thursday, October 29, 2026, 4:00–6:00 p.m. CDT (Austin local time), during the Work in Progress II & Demos session.
- Where: Shirley Bird Perry Ballroom, Texas Union, The University of Texas at Austin, 2308 Whitis Ave, Austin, TX 78712, USA.
Our presentation, Demo: MobileClaw: A Notification-Driven, Human-Gated Mobile LLM Agent, is demo 89 in the official demo listing. See the conference program and venue information for details.
- Notification-driven assistance. Background triage identifies notifications needing attention. Selected notifications trigger a bounded follow-up that reads relevant context and prepares a suggestion. For a photo request, actual local search results can appear as candidate cards in the request's conversation. The user previews and selects a photo, then explicitly requests sharing and confirms the image and recipient.
- Shared state on the phone. Foreground chat and background assistance use the same notification state, conversation history, memories and action records. Local document and image indexing supplies task context.
- Structured tool execution. Tools use notification actions, Android APIs and Android 16 AppFunctions exposed by cooperating apps. The included TaskBoard app makes the cross-app example reproducible.
- User control over actions. Outward actions require confirmation by default. Full Automation Mode can run eligible proposals automatically; privileged modifications, AppFunctions calls and selected-photo sending still require approval.
The prototype and its prompts are developed for English-language use. An AppFunctions target must expose compatible functions; available operations depend on the app, Android version and permissions. Connecting a different model service does not establish that it matches the demonstrated model's task performance.
- Follow installation and APK downloads, or
build from source. We recommend Android 16+ for MobileClaw;
Android 8.0+ is the minimum supported version. Supported architectures are
arm64-v8aandx86_64. The TaskBoard AppFunctions example requires Android 16+. - Open Settings > Model service, choose a provider, and enter your own model and connection details. Distributed APKs contain no team API key.
- Use Test connection, save the service, and grant notification access.
- Keep Full Automation Mode off and follow the quick start to try a synthetic notification.
The standard APK includes text, OCR and image-label search. Optional MobileCLIP weights for image/text similarity are available through the source-build instructions. The notification-to-photo walkthrough uses a known fixture filename so the candidate/selection flow can be tried without those optional weights. TaskBoard and Shizuku are optional for ordinary chat and notification use.
| Guide | Contents |
|---|---|
| Installation | Signed APKs, verification and installation |
| Quick start | First connection, permissions and a sample interaction |
| Model services | Gemini, OpenAI, custom endpoints, local inference and credential storage |
| Skills and tools | 11 Skills, 24 tools, permissions, confirmation rules and service requirements |
| Demo setup | Notification actions, document search, calendar and TaskBoard examples |
| Architecture | Agent loops, shared state, tools, search and action confirmation |
| Data and permissions | What stays on the phone, what reaches a model, and review/export controls |
| Building and testing | Toolchain, optional models, development builds and tests |
Boyuan Yang, Alexander Du, Ozzie Hassel, Jash Uganandan, and Wei Gao. Demo: MobileClaw: A Notification-Driven, Human-Gated Mobile LLM Agent. ACM MobiCom 2026. DOI: 10.1145/3795866.3848463.
@inproceedings{yang2026mobileclaw,
author = {Yang, Boyuan and Du, Alexander and Hassel, Ozzie and
Uganandan, Jash and Gao, Wei},
title = {Demo: {MobileClaw}: A Notification-Driven, Human-Gated Mobile {LLM} Agent},
booktitle = {Proceedings of the 32nd Annual International Conference on Mobile Computing and Networking},
year = {2026},
doi = {10.1145/3795866.3848463},
url = {https://doi.org/10.1145/3795866.3848463}
}Machine-readable citation information is in CITATION.cff.
app/ contains MobileClaw; demotarget/ contains the separate TaskBoard app;
tools/ contains fixtures, demo helpers, model downloads and evaluation tools;
docs/ contains the guides and presentation media.
Project code is MIT-licensed. See contributors and third-party notices for attribution. Third-party models, fonts and sample images retain their respective licenses.