What is AI Grid?

AI Grid (Grid) turns the computers you already own — Mac Studios, Linux GPU servers, idle workstations — into a private AI network. One machine hosts the grid router (no GPU needed), and other machines join as engines running models via llama.cpp, ComfyUI, or existing Ollama/vLLM/LM Studio instances. Apps send OpenAI-compatible requests to the grid URL, and Grid routes each request to the computer that has the right model. It supports text chat, vision, image/video generation, and experimental GRPO training across the fleet. A native desktop app (macOS/Linux) provides chat, schedules, and projects. The core is MIT-licensed; remote mode via Autonomous Relay lets machines behind NAT join. No accounts or cloud relay in local mode.

Should you self-host AI Grid?

AI Grid scores 48/100 on selfhost.directory (how scoring works) and ranks #59 of 125 in AI. It has 272 GitHub stars.

It shipped 10 stable releases in the past year. The latest, v0.3.61, came out on 27 October 2026. New versions land roughly every 1 days.

Setup is moderate: expect some configuration beyond a single command. It is MIT-licensed and written mostly in Python.

Its closest alternative on selfhost.directory is ComfyUI (93/100, 143 668 stars). It scores higher overall, so compare the two before you pick. See AI Grid vs ComfyUI side by side.

Features

Auto model selection
CLI for all operations
Distributed model routing
Distributed training (experimental)Marked experimental in docs; capture path and MLX backend described as young
Image and video generation
Model catalog with hardware-aware defaults
Multiple named grids
Native desktop app
No account or cloud relay (local mode)
OpenAI Compatible API
Remote mode via relay
Vision model support
Works with existing engines