Google AI Edge Gallery now helps the Mannequin Context Protocol (MCP), notifications reminders, and protracted chat historical past—offering builders with a showcase to construct linked, automated, on-device agentic experiences.
Google AI Edge Gallery (Android / iOS) is an on-device AI showcase app permitting customers and builders to work together and construct with Gemma and different open fashions. Final month, we launched the potential to deploy agentic workflows immediately on cellular units utilizing Gemma 4. At the moment, we’re increasing the core capabilities of the app to deliver extra linked, proactive, and protracted interactions. By supporting the open-source Mannequin Context Protocol (MCP), native notification reminders, and chat historical past, we’re making the app a extra helpful place to expertise and chat with fashions domestically, providing you with a playground to discover how one can coordinate duties throughout your knowledge and routines.
To enhance an LLMs on-device functionality, they want a standardized approach to work together with the world exterior the app sandbox. Based mostly on developer curiosity in additional versatile agent architectures, Google AI Edge Gallery now helps the open-source Model Context Protocol (MCP) over Streamable HTTP as an experimental function within the Android app. The replace to the iOS app is coming quickly.
By registering a valid MCP URL inside the app, the app dynamically imports device definitions and useful resource schemas immediately into the on-device mannequin’s system immediate. The reasoning and decision-making then occur solely in your cellphone. If you ask a query, Gemma 4 mechanically determines which device is required and generates the decision domestically; the request is then executed by the MCP server, which will be operating on your house pc or a safe cloud endpoint.
This structure permits the cellular gadget to coordinate advanced duties throughout varied knowledge sources and purposeful instruments. For instance:
- Productiveness increase: Hook up with Google Workspace MCP to let your cellular agent question your calendar for upcoming occasions or test your inbox for payments and tickets data.
- Navigation assist: Use the Google Maps MCP to ask about close by spots or journey instances utilizing pure language, making your agent a extra succesful journey companion.
- Internet Retrieval: Hook up with net fetch MCP to permit your agent to retrieve and parse content material from any URL. This permits the mannequin to purpose over real-time data—corresponding to information or documentation—offering fast summarization or solutions on the go.
Suggestions for builders:
As a result of on-device fashions function inside a smaller context window than giant server-side fashions, we advocate conserving your MCP device descriptions brief. When returning knowledge, sending again bite-sized snippets is most well-liked over lengthy textual content to maintain the expertise quick and responsive.
Technical documentation and instance configurations can be found in our GitHub repository.
Scheduling routines with notification reminders
Thus far, AI interactions within the app have been reactive—the consumer opens the app and kinds a immediate to start out a session. To allow extra automated use instances, corresponding to scheduled interactions, we’ve constructed a brand new “Schedule Notification” ability that makes it simple to arrange handbook routines.
In case you inform the agent, “Remind me to log my temper each night time at 10 PM,” it schedules an area notification. If you faucet that notification, the app opens on to the precise device and begins a session with Gemma 4, prepared to assist.
Listed here are just a few examples of how these proactive routines change the expertise:
- Temper monitoring and insights: A every day nudge asks the way you’re feeling. Over time, your agent can securely look again at your temper log historical past to identify tendencies and supply wellness insights.
- “Be taught one thing new”: A every day notification invitations you to be taught a brand new idea. Tapping it launches the agent to summarize a random or chosen subject right into a visually polished, shareable infographic card.
- Calendar briefings: Get a morning alert that summarizes your day. The mannequin shortly reads your native calendar and offers you a concise briefing in your schedule and any gaps in your day earlier than you even go away the home.
Extra app enhancements: Continuity and management
Along with new agent expertise, we’re addressing your high group suggestions by introducing every-day app enhancements that present extra flexibility when experimenting with on-device fashions.
Utilizing the quick prefill functionality of the LiteRT-LM backend, we’ve added assist for persistent chat historical past, permitting you to renew periods whereas sustaining the state of textual content, pictures, and audio inputs. On trendy cellphone GPUs, prefill speeds can exceed 3,000 tokens per second, permitting the mannequin to revive lengthy session contexts nearly immediately.
Alongside session continuity, we’ve launched the power to edit the customized system immediate immediately throughout the chat settings. This supplies builders with the management to experiment with immediate engineering strategies, outline particular mannequin personas, or implement strict output constraints for his or her on-device brokers.
The Google AI Edge Gallery ecosystem is constructed on an open-source toolkit. Whereas the talents we design present inspiration, the true energy of innovation resides inside our developer group.
On our official GitHub Skills Discussion page, the group is utilizing on-device fashions and edge {hardware} to construct highly-tailored, utility-focused workflows throughout many classes:
- Lengthen data base: Light-weight net search integrations that may present up-to-date data, corresponding to climate and forex change price.
- Agentic prompting: Educational constructing blocks that present context administration and permit fashions to work together in a extra customized manner based mostly on consumer preferences.
- Helpful parsers: Parsers that flip multimedia corresponding to picture, audio, and html into structured data, permitting customers to conduct semantic search on helpful private data, corresponding to expense or bookmarks.
- Enjoyable and studying: Card turbines that flip textual content into shareable visible infographics, quiz, language translator, offline puzzle video games, and even expertise designed for pets.
Obtain the newest model of the app from the Play Store and App Store, check out totally different expertise, and upvote your favorites! For builders—whether or not you’re bridging your knowledge ecosystem with customized MCP configurations, creating skills for various workflows, or constructing your personal app—head over to our repository, share your ideas, and tell us what you’d prefer to see subsequent within the app.
Acknowledgements
We would like to increase a particular because of our vital contributors for his or her work on this mission: Ashley Lin, Cormac Brick, Eric Yang, Geonsun Lee, Glenn Cameron, Hriday Chhabria, Ian Ballantyne, Jenn Lee, Matthias Grundmann, Sachin Kotwani, Na Li, Olivier Lacombe, Omar Sanseviero, Rishika Sinha
Innovation within the app expertise can be pushed by our developer group. We’re deeply appreciative for his or her work in growing and contributing new agent workflows and expertise: GitHub Discussions.
Discover this announcement and all Google I/O 2026 updates on io.google.