Local AI for the ideas, files and projects I carry on my iPhone.
Pocket Pi is a native app for working with downloaded AI models. Chats, documents and project materials live together, with phone inference as the default and explicit choices about what to download or share.
I submitted the first App Store candidate on 4 October, after the internal beta and native integration work. The product site brings together setup help and privacy information. I’m continuing to test phone performance and longer image-creation sessions.
Along the way
What’s changed, what I’m trying, and what comes next.
Pocket Pi has a first App Store submission, a dedicated product site and more useful examples of local work.
With the local assistant taking shape, I tried it with a small creative project: a story, a picture, a photo question and notes that belong together. The selected local models produced the text and photo answers. For the gallery, I rendered artwork locally on the Mac, then imported it into the app.
That exercise also found a useful small defect. A garden checklist could contain perfectly ordinary tasks such as selecting and planting, while the reminder extractor kept only some of them. I extended the supported action wording so those tasks can reach the review sheet. Saving still requires a separate decision; an answer does not silently become a system reminder.
Pocket Pi now has a dedicated product website, setup help and privacy pages. They explain the differences between chat, photo understanding and image creation, the downloads each needs, and the choices around optional connections. The point is to make the first useful step clearer before someone commits storage to a model.
I submitted version 1.0.0 build 17 for App Store review on 4 October. I still need to confirm its public App Store availability. The images here are from that release preparation; phone speed, memory, heat and longer image-creation sessions are still part of the testing ahead.
A new native iPhone app brings local chat, documents, projects and image creation into one place.
The Mac setup taught me a lot about running models locally. Pocket Pi takes that interest onto an iPhone: a native app where I can ask a question, bring in a document, keep related work in a project and return to it later. Models are downloaded explicitly, and phone inference is the default.
This week brought the pieces together. Projects can hold materials, instructions and saved answers; the library can take text, files, PDFs and scans. Chat and photo understanding use llama.cpp, while image creation has a separate MLX engine. I have been working on the less glamorous parts too: recovering interrupted downloads, keeping drafts attached to the right conversation and preserving the original description when an image request needs shortening.
The latest build also makes it feel more at home on the phone. Share intake waits for review, widgets and Shortcuts lead into useful actions, and generated images offer Photos saving and separate Markup copies. A completed answer can become a reviewed Reminders list. Translation, optional title-only Spotlight search and Live Activity progress round out the native integrations.
Version 0.9.0 build 14 is in internal TestFlight, with a successful installation confirmed on 3 October. The main flows have been through simulator checks, but I still need to test sustained performance, image quality and accessibility on a physical phone. For these simulator screenshots, I used an image rendered on the Mac to try the image controls.