VARIANT-1
0.1.1 previewDownload
A LIVING AI WORKSPACE

Intelligence.
In its element.

A free desktop app where AI does real work:
running code, handling files, browsing the web.
Each chat keeps its results alive for the next step.

Download VARIANT-1 See how it works

Free & open source · Windows, macOS, Linux

V—001 / MUTABLE MATTER
GENERATIVE STUDY 001
BUILT FOR THE WORK BEYOND THE PROMPT.LOCAL OR CLOUD YOUR MODEL

01 THE WORKSPACE

A conversation is
only the beginning.

Ask for something and the model gets to work: it writes a little Python, opens your files, checks a website, and shows you every step. Files, a browser, a terminal, and a Review pane sit right beside the chat, or in windows of their own.

From a messy folder
to something that works.

Tidy a spreadsheet, explore a codebase, or write a script. The model runs it, checks the output, and keeps the working data ready for your next request.

FILES → CODE → CHECKED OUTPUT
VARIANT-1 / MAIN DECKSCREEN RECORDING · OCTOBER 2026
One workspace. Every panel can become its own window.Sample project · chat staged for the recording

Recorded in the current preview with a sample project. The conversation was set up for the recording; the panels and windows are the app itself. The interface will keep changing before a stable release.

02 LIVING STATE

Your work stays alive
between steps.

Every chat grows its own Python session. The data and helpers the model creates live there, so a follow-up builds on them instead of starting over. Try it.

CULTURE / ONE CHAT’S PYTHON SESSIONSESSION 01 · LIVE
ILLUSTRATION, NOT THE APP INTERFACE

TRY IT · THREE STEPS

    What lasts, and for how long

    03 THE METHOD

    Keep the state.
    Change the approach.

    The model works through a task in Python, combines tools as it goes, and can even improve those tools along the way.

    STRUCTURE / VARIATION
    ONE SYSTEM. MANY POSSIBILITIES.
    01 /

    Build on
    what’s already there.

    Each chat’s live Python session keeps data, imports, and helpers between steps, ready for the next request.

    02 /

    Tools that work
    together.

    Code connects the steps: read a file, reshape the data, open a page, check the result. Useful steps become reusable helpers.

    03 /

    Room to adapt.
    You stay in the loop.

    Turn on mutation and the model can propose better versions of its tools, test them, and roll them back. You can steer or stop it at any time.

    ADAPT THE METHOD. KEEP THE USEFUL WORK.Mutation is optional · Off by default

    04 MODEL CHOICE

    Your model.
    Your call.

    Run one on your own machine, use an API key,
    or sign in with a plan you already have.

    Compare model connections
    Same workspace. The model runs on your machine.

    Run locally

    Point VARIANT-1 at a GGUF model and it runs llama.cpp for you, or connect LM Studio. No per-message fees.

    LOCAL MODELS

    Use an API key

    OpenAI, Anthropic, Google Gemini, xAI, OpenRouter, DeepSeek, and dozens more, plus any OpenAI-compatible endpoint. Your provider bills you directly.

    API KEYS + CUSTOM ENDPOINTS

    Sign in with a plan

    Use a ChatGPT subscription, or a desktop service you’re already signed in to, such as Ollama.

    ACCOUNTS + SUBSCRIPTIONS

    05 BEYOND ONE CHAT

    Work that
    keeps going.

    Give the agent a bigger objective, let helper agents share the load, and put routine jobs on a schedule.

    How goals and automations work
    START → STEPS → GOAL
    01 / GOALS

    Bigger objectives

    Start a Goal and the agent plans the steps, keeps progress notes, and picks up where it stopped.

    02 / AGENT TEAM

    Helpers on call

    Big jobs can be split between helper agents, each with its own session. You can see what every helper is doing.

    03 / AUTOMATIONS

    Routines that run themselves

    Run an instruction on a schedule or when something happens on your computer, like a morning briefing.

    06 GOOD QUESTIONS

    A few things
    worth knowing.

    PATTERNS / POSSIBILITIES
    Is VARIANT-1 free?
    Yes. It’s free and open source under the MIT License, with no account or subscription. You bring the model: one running on your own computer, or a provider you pay directly. View the source on GitHub
    Which computers does it run on?
    Windows 10 and 11 (64-bit), Macs with Apple silicon, and 64-bit Linux. Each download on the download page comes with a checksum and notes on what was tested.
    Does my data stay on my computer?
    The app and its Python sessions run on your computer. If you use a cloud model, the messages and results it needs are sent to that provider. With a local model they stay on your machine, though browsing still uses the internet.
    What is a “live session”?
    Each chat has its own Python session that keeps running between messages, so data and helpers from earlier steps are still there. A restart clears it, unless you turn on checkpoints (off by default), which can save supported values.
    What does mutation do?
    It lets the model improve its own tools within a chat: propose a change, test it, switch it on, and roll it back if needed. It’s off by default and never retrains the model.
    Is it safe to let it act on my computer?
    VARIANT-1 isn’t a sandbox: it acts with your own permissions. Start with copies of important files, give clear instructions, and use Stop when you need to. Stopping can’t undo something that already happened. How you stay in control
    FREE PUBLIC PREVIEWWINDOWS · MACOS · LINUX

    Your next idea.
    A place to work on it.

    Bring a model.
    See what you can make together.

    Download VARIANT-1 Read the setup guide
    FREE APP. BRING YOUR OWN MODEL.
    VARIANT-1 with four separate windows arranged around the chat: Files at the top left, Review at the top right, the Log monitor at the bottom left, and Overview with CPU, memory, disk, and GPU graphs at the bottom right.

    VARIANT-1 preview, October 2026, with a sample project. The conversation was set up for this picture; the windows are the app itself.