AX Desktop

Overview

Three ideas explain every design decision in AX:

Local first, local only. Inference happens on your machine. AX makes no network calls except the ones you ask for: downloading a model, fetching a page for a tool you enabled. Offline is not a degraded mode; it is the default mode.

Your data is a folder. Conversations, settings, and agent state live in plain files under one directory (~/ax by default). Copy it to back it up. Delete it to be forgotten. Read it with any editor.

Standard interfaces. Models load from the GGUF format the open ecosystem already uses, and the built-in server speaks the OpenAI-compatible dialect that most tooling already targets. AX would rather join the ecosystem than invent one.

Install

Windows

Run the installer and launch AX from the Start menu. NVIDIA GPUs are used automatically when CUDA drivers are present; otherwise AX runs on CPU with AVX2 acceleration.

macOS

Drag AX into Applications. On Apple Silicon, inference runs on the GPU through Metal and shares unified memory, so 16 GB machines comfortably run 8B-class models.

Linux

Use the AppImage (make it executable and run it) or the Flatpak. CUDA, ROCm, and Vulkan backends are selected automatically and can be overridden in Settings → Engine.

Quickstart

First launch takes about two minutes:

1. AX profiles your hardware and recommends a starter model that fits your memory. Accept it or pick from the model library.

2. The model downloads with a visible license notice from its publisher. Weights are stored once in ~/ax/models and shared across every conversation.

3. Type. Everything after this point works with the network unplugged.

Models

The library curates open-weight models by what they are actually good at: everyday chat, deep reasoning, code, vision, and speech. Each entry lists its memory footprint per quantization so you know what fits before you download.

Anything not in the library still works: Import GGUF accepts any local file, and a Hugging Face URL pastes straight into the search box.

Details and the current list live on the models page.

Local API server

Toggle Settings → Server and AX exposes your loaded models at http://localhost:4891/v1 with OpenAI-compatible routes (/chat/completions, /completions, /embeddings, /models).

curl http://localhost:4891/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ax/current",
    "messages": [{"role": "user", "content": "Hello from my own hardware."}]
  }'

The server binds to localhost only unless you explicitly change it, and it can require a bearer token you set yourself.

Agents & tools

An agent in AX is a model plus a set of tools plus a standing instruction, saved as a file you can version. Tools ship in the box for file access, shell commands, and web fetch; each one is off until you grant it, per workspace, with the grant visible in the sidebar the whole time it is active.

Agents run three ways: on demand from the chat, on a schedule, or watching a folder. Every run writes a plain-text transcript to ~/ax/runs so you can audit exactly what a tool did and why.

Extensions

An extension is a directory with a manifest and an entry point, running out of process with only the permissions its manifest declares. If you can write a few hundred lines of TypeScript or Python, you can teach AX something new and distribute it from any git repo; there is no store to apply to.

Privacy

AX sends no telemetry, no crash reports, no usage pings. There is no account system. The only network traffic AX ever initiates is a model download you clicked and whatever a tool you explicitly enabled was asked to fetch. This is verifiable in the source, and the source is the product.