AriCore runs large language models on your phone. You bring the model file, it runs locally. Nothing leaves the device. If you also use aricode/aridesk on your desktop, you can pair the two and chat from your phone while the desktop does the heavy lifting. The pairing is end-to-end encrypted — the relay server only ever sees ciphertext, and neither end stores messages on anyone else's computer. What you can do with it: Load a GGUF file and chat offline. Llama, Qwen, Mistral, Gemma, whatever you can fit in memory. Quantized 3B–7B models work well on recent iPhones. Bigger models work too if the device has the RAM for it. Connect your own API keys for OpenAI, Anthropic, or Google if you want frontier models without giving your data to a third-party chat app. Keys stay in the iOS Keychain. Pair with aricode desktop to use your existing sessions, approve tool calls, and keep going from the couch. Handy if you run long jobs on a workstation and want to glance at them from the kitchen. Build a self-paced course on anything. The Study tab spins a curriculum from a topic and a goal — lesson plus quiz, with optional cloze, written-answer, and code exercises per session, plus a midpoint and final exam. Tutor chats stay walled off from your personal context — no tool access, no memory writes, no calendar lookups. Finish a course and you get a certificate and a one-tap PDF export of the entire workbook, answers and grades included. Let the model reflect overnight. Dreams (optional) read recent conversations while the device is charging and surface proposals — memories worth keeping, custom skills you'd actually use, persona suggestions, watchers for things you care about. Every proposal is presented for review; nothing applies without your tap. Skills are local tools the model can call — plain markdown plus a small JSON config. Build them by hand or have the model draft one from a description. The skill creator runs in an isolated context so your personal data never leaks into the generated code. Voice mode runs entirely on-device. On-device speech recognition for input, Kokoro TTS for output, no cloud round-trip and no microphone audio ever leaving the phone. Why it exists: Most mobile AI apps are thin clients for somebody's backend. Your prompts sit on their servers, get mined for training, get tied to an account. This one doesn't do any of that. No account. No analytics. No crash reporting that phones home. The code path for "talk to a model" goes from your screen straight to the model — either on this device or over an encrypted tunnel to your own desktop. A few things worth knowing: You provide the models. There's no in-app store or downloader. Drop a .gguf into the Files app under AriCore and it shows up in the model picker. Running a 7B model takes 3–5 seconds to first token on an iPhone 15 Pro. Smaller models are faster. API keys live in the Keychain. The app never sees them after you paste them in unless you re-open the settings. They're not synced anywhere. The desktop pairing requires aricode running on the same network or reachable over Tailscale. Setup is one QR code scan.
