off-grid-ai/OGAD

Private, on-device AI. Your models, your data — no cloud, no accounts, no API keys.
A local-first AI runtime + studio — run open models ( text, vision, image, voice, speech )
entirely on your machine, behind one OpenAI-compatible gateway. Plus an always-on layer that
sees, remembers, reflects, and acts — all on-device.
Download (macOS · Windows · Linux) ·
Features ·
getoffgridai.co ·
Get Pro ·
Slack community
Off Grid AI Mobile — the same on-device AI, on your phone  · 
 ·  100k+ downloads
---
What it is
Off Grid AI is a **local-first AI runtime** for your desktop. Download open models from the
built-in catalog (or any GGUF from Hugging Face) and use them across every modality — all
inference runs on your hardware via bundled llama.cpp, stable-diffusion.cpp,
whisper.cpp, and Kokoro. Nothing routes through a server we own; your conversations,
files, and models never leave your device.
Three things in one app:
1. **A studio** — chat (text + vision + reasoning), on-device image generation, voice
in/out, live artifacts/canvas, projects with RAG, and in-chat tools — a local
Claude/LM-Studio/Ollama with everything on-device.
2. **A gateway** — one local OpenAI-compatible API (http://127.0.0.1:7878/v1, no key)
for chat, vision, image, audio, and embeddings. Run it headless as just the gateway.
3. **Off Grid AI Pro** — an always-on private layer that _sees_ your work (screen → OCR),
_remembers_ it, helps you _reflect_, and _acts_ with your approval. On-device, opt-in.
A look inside
Chat — text, vision, reasoning, artifacts, on-device
Models — curated catalog + Hugging Face search
The Gateway — one local OpenAI-compatible API
Projects — group chats, RAG over your docs
Connectors (MCP) — add servers, use them in chat
Artifacts — HTML, React, SVG & Mermaid in a local sandbox
Off Grid AI Pro — the sees/remembers/reflects/acts layer
Private by default — runs on your machine, no account
Features (free & open source)
The free, open app is a complete on-device AI studio:
- **Chat** — text + vision, streaming, with a reasoning ("thinking") mode and per-chat
model settings (temperature, context window).
- **Image generation** — text→image and image→image via stable-diffusion.cpp (Metal).
Ships **SDXL-Lightning** (few-step, fast), SDXL, SD 1.5/2.1, and **Z-Image-Turbo**
(2026 flagship, ~8-step). Live per-step preview, progress + ETA, cancel, lightbox, and
an artifacts gallery of everything you've generated.
- **Voice** — speech-to-text (whisper) and text-to-speech (**Kokoro-82M**, multilingual),
plus a hands-free voice mode.
- **Artifacts / canvas** — the model's HTML, React/JSX, SVG, Mermaid, and Markdown render
live in a sandboxed iframe (no network/file access); Code/Preview toggle, download,
saved per chat & project.
- **Projects** — group chats, upload documents (txt/md/PDF/DOCX, image, audio, video) and
chat grounded in them (RAG with cited sources); per-project instructions.
- **Tools in chat** — an agentic loop calls local tools mid-conversation: built-ins
(calculator, datetime) plus any MCP connector you've added.
- **Connectors (MCP)** — add Model Context Protocol servers (none / token / OAuth) and use
them right inside chat. Preset catalog included.
- **Model catalog** — curated, size-bucketed recommendations + direct Hugging Face search;
download, manage, and set the active model per modality.
- **The Gateway** — one OpenAI-compatible endpoint for everything; see below.
- **Auto-update** — signed releases update themselves.
A full breakdown is in docs/FEATURES.md.
The Gateway
One local server (http://127.0.0.1:7878) speaks the OpenAI API:
| Capability | Endpoint |
| -------------------- | -------------------------------------------- |
| Chat (text + vision) | POST /v1/chat/completions |
| Text → Image | POST /v1/images (/generations, /edits) |