Deploy vs Token

On device private AI agents that live entirely on your iPhone. No accounts. No cloud. No tracking. Every conversation, every image, every document stays on your device — even with airplane mode on. Built for people who want a private ChatGPT or Claude alternative that works offline — without the account, the cloud round-trips, or the data harvesting. Pick an agent and start. Or build your own in five guided steps. — THE FIVE BUILT-IN AGENTS — Scout — your research agent. Searches the web with citations, summarizes findings, and remembers context across questions. Lens — your vision agent. Drop in a photo, screenshot, or whiteboard shot and ask anything about it. Ghost — your writing agent. Drafts, rewrites, edits, and shifts tone. The output reads like you, not like a chatbot. Vault — your document agent. Load PDFs, text files, or Markdown notes and ask questions. Answers cite the page they came from. Spark — your brainstorming agent. Pressure-tests ideas, runs lateral riffs, and breaks creative blocks. — BUILD YOUR OWN AGENT — A five-step guided wizard turns an idea into a working agent in under a minute. Choose its purpose, give it tools, pick a model, set its look. The on-device model writes its name, picks an emoji, and drafts the system prompt for you to approve. Unlimited custom agents. Each one keeps its own personality, its own toolset, and its own accent color. — TRUE ON-DEVICE AI — Powered by Apple's MLX framework with 4-bit quantized models that download once and then run forever, offline. Supported models include Qwen 3.5 in 0.8B, 2B, and 4B sizes — all multimodal with vision — plus Ministral 3B for fast text-only chat. More models are being added regularly, and we encourage you to contact us for requests. Pick the model that fits your iPhone. Switch any time. — FEATURES — · Streaming token output, native SwiftUI, dark-mode first · Per-agent accent colors and emoji · Extended thinking mode you can toggle per message · Web search via DuckDuckGo with proper citations · Native image understanding for screenshots, photos, and scans · Local document Q&A across PDF, TXT, and Markdown · VoiceOver, Dynamic Type, and Reduce Motion all fully supported — PRIVACY, FOR REAL — · 100% on-device inference. No model call ever leaves the iPhone. · No account. No sign-in. No email. · No analytics. No telemetry. No third-party SDKs reading anything. · Conversations, attached images, and uploaded documents stay on your device. · The only network call the app ever makes is your explicit web search, and only when you tap Scout. The privacy posture is a feature, not a marketing line. Your iPhone is the entire system. — REQUIREMENTS — iPhone running iOS 18 or later. Models download on first use; sizes range roughly from 600 MB to 3 GB depending on which model you pick. — SUPPORT — Questions, suggestions, bug reports: see the in-app Help screen.

Token AI is a private AI studio for your iPhone. Download open models, run them on the device, and keep every conversation on your iPhone. FRONTIER MODELS ON YOUR IPHONE • Experimental 35-billion-parameter mixture-of-experts models, Edge0 35B and MiniCPM-V 4.7 35B, stream their experts from storage. They need a large download, and MiniCPM-V 35B needs an iPhone with 12 GB of memory. • Ternary Bonsai 2 9B and Bonsai 27B: ternary and 1-bit models that fit more capability into iPhone memory, with text and image understanding. • Qwen3.5, Qwen3, Llama 3, Gemma, Phi-3.5, LFM2.5, Nemotron and Granite, in MLX, GGUF and Apple Core AI formats. • Import your own compatible MLX or GGUF models from Files. CHAT, REASON AND CODE • Reasoning models show their thinking, and long chats are compacted to stay within the model's context window. • Attach photos and files in any chat. Vision models read images directly; other models get a description from an on-device visual model. • Choose a model for each conversation, pin chats, and search every message. • Optional web search, a document knowledge base and GitHub repository tools. LENS, VOICE AND IMAGES • Lens: use the camera or a photo and ask about it with SmolVLM 2, Gemma 3 and other vision models. • Voice conversations with on-device speech recognition (Whisper) and natural voices (Kokoro, Kitten TTS). • Image Studio: create pictures with Stable Diffusion, SDXL Turbo and other local image models, with optional PIN-protected parental controls. FOR BUILDERS • An authenticated local API server for OpenAI-, Anthropic- and Ollama-compatible tools on your trusted network. • Optional API providers with your own keys. • An on-device classifier workbench, Shortcuts and Siri actions, a share extension, and Live Activities for downloads. PRIVATE BY DESIGN No account, analytics or ads. Conversations are stored only on your iPhone. Model downloads, web search, API providers and GitHub use the network only when you turn them on. Requires iOS 27 or later. Works on iPhone Duo. Available models depend on your iPhone's memory, and model files are downloaded separately and can need several gigabytes of storage. AI results can be inaccurate; review them before relying on them.