First run and startup

The first time you open AgenC Desktop, it prepares AgenC Core, the agent runtime inside the app, and downloads the speech model for voice input. Then it helps you choose where your model comes from. It asks no questions until then, and it does not use the microphone.

First run

The first run opens in a small window and goes through three screens:

1. Welcome: the AgenC mark and a handwritten welcome, for a few seconds.
2. Preparing: "preparing agenc", with a progress bar. The caption follows what Core is doing: "getting things ready", "setting up the essentials", "preparing your workspace" and so on, until "you're all set".
3. Ready: "AgenC Ready". Then the window grows to its normal size and the app opens.

While Core prepares, the app downloads Whisper's Base speech model, about 141 MB, so you can talk to the agent later. The caption says "downloading the speech model" and shows how many MB have arrived. Setup finishes only when both Core and the speech model are ready, so the first run needs an internet connection.

If something fails, the screen shows what happened and a retry setup button. A download that stops moving shows "download stalled, still waiting" before it gives up.

The first run only moves on while its window is visible and focused. If you quit halfway, it starts again at the next launch. If you install over an older version of the app, its data is found and the first run is skipped.

Choose how to get a model

After setup, the home screen offers three cards, once:

• Local model ("hardware matched"): finds, downloads and starts a model that fits this computer. No account, no key.
• AgenC model credits ("check your free access"): signs you in to your AgenC account in the browser, to use AgenC's hosted models.
• Your own API key ("bring your provider"): opens Settings > Providers, where you paste a key or sign in with X or ChatGPT.

The cards come back at each launch until one choice completes. You can connect more providers at any time; see Providers and models.

The local model option

1. The chat says "find the best model for my hardware" and reads your hardware: processor, usable memory and graphics.
2. A sheet asks "Which one should this machine run?" Each model shows its download size, its likely speed on this computer, its quantization, and "chat only, no tools" when it cannot use tools. None of these downloads nothing.
3. If nothing on the computer can run the model, the app offers to install Ollama: where Homebrew is installed it installs Ollama for you, otherwise it opens Ollama's download page. Not now installs nothing.
4. The app downloads the model and shows its progress, then says the model "is installed and running on this machine. It is now the model this app uses."

The recommendation needs an internet connection, or the llm-checker tool installed locally. If no model fits the computer's memory, the app says so and installs nothing. While this runs, Retry setup and Back to setup sit above the composer.

The loading screen

From the second launch on, the app opens behind a loading screen while Core starts: the AgenC mark over a blurred window, with one short phrase saying what it waits for, such as "starting agenc", "loading your sessions", "loading plugins" or "almost ready". It fades out once your sessions, providers, plugins and skills are loaded.

If Core fails to start, the screen shows the reason and two buttons:

• Try again reconnects, and starts the app's Core if it is not running.
• Open anyway removes the screen for this window, so you can reach Settings.

The two buttons also appear when loading takes longer than a minute.

Close and quit

• macOS: closing the window keeps AgenC and Core running, so routines and remote connections carry on. Click the Dock icon to open a window again. Quit with AgenC > Quit AgenC.
• Windows and Linux: closing the last window quits the app.

Quitting stops everything the app started: terminal tabs, the browser, Core with its agents, and remote connections. The windows close at once, and the app waits for Core to finish closing its sessions before it exits, so opening it again right away works. It does not ask for confirmation, even while a task runs. A Core that was already running before the app started is left running.

If remote connections are on when you close the window, the app asks first: "Keep AgenC running for your remote devices?", with Keep running (minimizes the window), Stop connections and close, or Cancel.

Opening AgenC while it already runs brings the existing window to the front instead of starting a second copy.