Your AI. Your hardware. Nobody else's business.

AX is an open-source replacement for cloud chatbots. It ships with curated open-weight foundation models and an application ecosystem, and every token is generated on your own machine. Nothing you type ever leaves the room.

Get early access Read the docs

Open source. Apache-2.0. Offline by default. No account, no telemetry, no exceptions.

Everything a cloud chatbot does. Nothing it takes.

Runs entirely on your hardware

AX detects your GPU, CPU, and memory, then picks quantizations that actually fit. Apple Silicon, NVIDIA, AMD, or plain CPU. Pull the network cable out mid-conversation and nothing changes.

Foundation models built in

A curated library of open-weight models, one click to download, swap mid-chat. Chat, reasoning, code, vision, and speech. Bring any GGUF you like.

An ecosystem, not an app

Agents, tools, and extensions run against a local OpenAI-compatible API. Anything that speaks that dialect can plug straight into AX.

What ships in the box.

Chat that stays home

Full conversation history, files, and search stored in plain local files you can read, back up, and delete. Your history is a folder, not a subpoena target.

Local API server

One toggle exposes your models on localhost as an OpenAI-compatible endpoint. Point your editor, scripts, and agents at it and cancel the metered key.

Agents and tools

Give models hands: file access, shell, web fetch, and custom tools, each behind an explicit permission you grant per workspace. Agents run on a schedule or on demand.

Extensions

An open extension format for connecting AX to the rest of your machine. Write one with a manifest and a few hundred lines; ship it without asking us.

Get AX.

First public build is being hardened now. Early access builds go out by email while we finish the public release.

Windows

Windows 10/11, x64. NVIDIA acceleration via CUDA, CPU fallback everywhere else. Request build

macOS

Apple Silicon, macOS 14+. Metal acceleration, unified memory aware. Request build

Linux

x64 AppImage and Flatpak. CUDA, ROCm, and Vulkan backends. Request build

Every build is reproducible from source. When the public release lands, binaries and the full repository publish together; see the changelog for where the release train currently is.

The fine print, in plain sight.

AX is developed by Agentix Labs, LLC and licensed under Apache-2.0. Model weights are downloaded from their original publishers and remain under their own licenses, including the Meta Llama Community License, Apache-2.0, MIT, and the Gemma Terms of Use; AX shows you the license before the first download of every model and never rehosts weights whose terms forbid it. AX is an independent project. It is not affiliated with, endorsed by, or sponsored by OpenAI, Anthropic, Google, Meta, or any model publisher; all product names and marks belong to their respective owners. AX collects no analytics and sends no telemetry. There is nothing to opt out of.