Opening GitHub…Opening Discord…

    Atlarix 15.19.3: A Model for Each Task, and Switch to Build in One Click

    November 2026~3 min readProduct

    Atlarix 15.19.3 is out today. Atlarix is an AI coding agent by Norah Labs that plans, edits, runs and reviews code in your own projects, and it asks before it changes a file. This release gives you more say in how it works: which model does which part of the job, how a read-only mode gets to Build, and which calls your own rules approve.

    A model for each task

    Open Integrations → Models & providers in the app. A new section at the top, Model for each task, has a row for Explore, Plan, Build and Sub-agents. Pick a model for any of them, or leave a row on Same as chat to use the chat's model. Atlarix Auto and Zenith keep choosing their own.

    So you can plan with a model that reasons well and build with a faster one, or keep research on a local model. If a model you chose can no longer run, for example because its key was removed, its row tells you why, and the task stops with a message saying where to change it. It never switches to a different model without telling you.

    On a free account, a task can use the models from your own keys. The CLI and the editors get the same setting in a later release.

    Switch to Build in one click

    Explore and Plan don't edit files. Ask for a change in one of them and Atlarix now shows a card: This needs Build mode. Switch to Build? Press Switch to Build and it carries on with your request. Press Stay and nothing changes.

    Before, you got a question in plain text and had to switch modes yourself. In runs where nobody can answer a card, such as automations, the card doesn't appear.

    Your edicts can approve a call

    An edict is a rule you give Atlarix about what it may do. A pre_tool_use edict can now answer allow, and the approval card for that call is skipped. It works like a Claude Code hook: an edict written for Claude Code that returns permissionDecision: "allow" works here too.

    When several edicts match, the strictest answer wins: deny, then ask, then allow. Your deny rules still win over an allow, and anything that reaches outside your project, such as sudo, an install or a dangerous command, still asks you.

    The headless CLI is on npm

    The headless CLI runs the agent from the command line, for CI and scripted runs. Install it with npm i -g atlarix on macOS, Linux or Windows, then run a task with atlarix run. See the headless CLI guide.

    Update from the app, or download Atlarix at atlarix.dev. The full list of changes is in the changelog.

    Questions

    Can I use a different model for planning and building in Atlarix?
    Yes. In the app, open Integrations → Models & providers and choose a model for Explore, Plan, Build and Sub-agents under Model for each task. Unset tasks use the chat's model.
    What happens when I ask for a change in Explore or Plan?
    Atlarix shows a card asking whether to switch to Build. Switch and it carries on with your request; stay and nothing changes.
    Do Claude Code hooks work as Atlarix edicts?
    Yes. A pre_tool_use hook that returns permissionDecision "allow" skips the approval card in Atlarix too. Deny rules and anything outside your project still ask.