Off Grid AI

The Swiss Army Knife of On-Device AI

Chat. Generate images. Use tools. See. Listen. All on your phone or Mac. All offline. Zero data leaves your device.

---

BUILT BY

---

Off Grid AI Pro

A voice, personas, and actions. $119 for life, or $7.99/month.

Pro is an optional, additive tier. It gives the assistant a voice that talks back, personas you shape, and the tools to draft real actions you approve. One license covers your phone and your Mac. All on-device.

What Pro adds

• Voice mode - the free app transcribes your speech; Pro adds on-device Kokoro text-to-speech, so it talks back and you run the whole thing hands-free. The voice runs in your phone's RAM.
• Custom personas - give each assistant its own system prompt, voice, and persistent memory, so it stays in character across conversations.
• Draft, then approve - connect Calendar, email, and MCP servers like Linear, Notion, and GitHub. It drafts the reply or files the ticket and waits. Nothing sends without your tap.
• Personal Mesh sync - your phone and your Mac share chats, files, and settings over your own network. There is no relay. See the Personal Mesh guide.

→ Get Pro access - $119 once and it is yours forever (the price climbs as more people join, never down), or $7.99/month.

Pair it with Off Grid AI Desktop on your Mac. One Pro license covers both.

---

Not just another chat app

Most "local LLM" apps give you a text chatbot and call it a day. Off Grid AI is a complete offline AI suite — text generation, image generation, vision AI, voice transcription, tool calling, and document analysis, all running natively on your phone's or Mac's hardware.

---

What can it do?

Text Generation — Run Qwen 3, Llama 3.2, Gemma 3, Phi-4, and any GGUF model. Streaming responses, thinking mode, markdown rendering, 15-30 tok/s on flagship devices. Bring your own .gguf files too.

GPU & NPU Acceleration — Your phone has silicon sitting idle. Off Grid uses it. Adreno GPUs via OpenCL run 20-40 tok/s on a Snapdragon 8 Gen 2+, against 15-30 on CPU; Apple Silicon uses Metal. The app detects what your device has and defaults to the fastest backend that works, and you can override it in Settings. The Hexagon NPU (Snapdragon) is there too, marked experimental because it is — it only accelerates Q4_0 and Q8_0 quants, a K-quant silently falls back to CPU, and some model architectures come out garbled on it. Models that can actually use the GPU or NPU are badged in the model list, so you pick the right quant before you download 4GB.

Remote LLM Servers — Connect to any OpenAI-compatible server on your local network (Ollama, LM Studio, LocalAI). Discover models automatically, stream responses via SSE, store API keys securely in the system keychain. Switch seamlessly between local and remote models.