Atlarix Auto: A Free AI Coding Agent That Picks the Model for Each Step

Atlarix is a desktop coding agent: it plans, edits your code, runs commands in a per-OS sandbox, and runs your tests before it calls anything done, with every change waiting for your approval. Auto is the model setting that decides which language model does each piece of that work. You do not pick a model, you do not paste an API key, and you do not need to know which models exist. You pick Auto and give it a task.

A coding session is not one kind of work. Reading a codebase to answer a question, writing a change, planning a refactor and reviewing a diff need very different amounts of reasoning — and paying frontier prices for all of them is waste, while using a small model for all of them is a false economy. Auto routes on what each request is:

The routing happens per request, not per session, so a long task that suddenly needs to look at a screenshot gets a vision-capable model for that one step and returns to the cheaper model afterwards. If a step fails, Atlarix retries it; if you want more muscle on a hard turn, a “Try with a stronger model” button reruns it one level up. Model names are deliberately hidden under Auto: the composer shows “Auto”, and what it cost.

On an account with no credit, Auto in the desktop app costs nothing. It runs on OpenAI models through OpenAI's data-sharing program: OpenAI provides free daily capacity for traffic that it may use to improve its models. That is the whole reason it is free, and we would rather say so plainly than bury it. Before your first free message, Atlarix shows a one-time notice explaining this, and you choose whether to go ahead. If you do not want your prompts shared, do not use free Auto for that work — every other path in Atlarix is private (see below). Free Auto is available in the desktop app, and to Atlarix Reviewer for public repositories only; the Chrome extension runs on paid Auto.

Free capacity is shared, so it is divided fairly rather than first-come, first-served. Each account gets a daily share sized from what is left and how many people are using it that day, with a floor so everyone gets real work done and a ceiling so no single account drains it. When your share runs out, Auto falls back to a free model rather than cutting you off; only past that does it pause until the daily reset — and the message tells you the reset time on your own clock, not in UTC. Atlarix also counts every running request against the pool before it starts, so free usage can never run past the free allowance and quietly turn into a bill.

Privacy is the reason Atlarix exists, so the free trade is opt-in and avoidable. With Atlarix Pro or credit, Auto is paid and private: nothing it handles is shared or used to train models. Your own API keys work exactly as before, billed to you by your provider. Local models through Ollama or LM Studio never leave your machine at all. And Atlarix itself never trains on your code, in any mode.

The Atlarix app is free — every feature, no gating. Paid inference comes two ways: Atlarix Pro at $19/month, which includes Atlarix Core usage across the desktop app, the Chrome extension and Atlarix Reviewer, or pay-as-you-go credit where you pay only for what you use and see the cost of every reply. Auto on a paid account holds only what a step could cost on the model it expects to use, charges the real cost when the step finishes, and never downgrades your model because a balance is getting low — a step that starts, finishes.

Download Atlarix from atlarix.dev (it updates itself if you already have it). New installs start on Auto; if you picked a model before, switch to Auto in the model picker. Then give it a real task on your own code — that is the only benchmark that matters.