# Getting Started

Actual runs models on computers you control. Install and authorize a computer, start Actual, download a model, then load it before using Chat or the API.

## Compatibility

| Platform | Architecture | GPU backend |
| --- | --- | --- |
| macOS | Apple Silicon (M1 or later, ARM64) | Metal |
| Linux | x86_64 | NVIDIA CUDA or AMD ROCm |
| Linux | ARM64, including DGX Spark | NVIDIA CUDA |
| Windows | x86_64 | NVIDIA CUDA or AMD ROCm |

Intel Macs and Windows ARM64 are not supported by the public builds. Check the platform guide for the required tools and GPU runtime. A published build does not mean every model or combination of computers supports distributed inference.

## First working model

1. Sign in to [actual.inc](https://actual.inc), or [create an account](/register).
2. Open [Computers](/console/computers) and choose **Add lead computer**.
3. Copy the command for your platform and run it on that computer within 10 minutes. Treat the command as a credential.
4. Open a fresh terminal, run `actual --version`, then `actual whoami`. If login is incomplete, run `actual login` and open the exact device URL it prints. A valid code is authorized automatically after sign-in.
5. Run `actual` to start the background daemon, then `actual status`. Confirm the computer appears online in Computers.
6. Choose a model from [Models](/models) or the [registry](https://models.actual.inc), download it, and wait for installation to finish. A fresh installation may have no models.
7. Run `actual models list`, then `actual models load` to choose an installed model. Wait until `actual status` reports the expected loaded model.
8. Open [Console > Chat](/console/chat), select the cluster and model, and send a message. Chat uses your signed-in session.
9. For an external application, create a credential in [User > Keys](/user/keys) and follow [API Quickstart](/docs/api-quickstart).

See [Model Discovery](/docs/model-discovery) for download commands, progress, cancellation, and loading from the website.

## Choose your installation guide

- [Machine Installation](/docs/machine-installation) explains credentials, startup, and recovery.
- [Set Up on macOS](/docs/macos-setup) covers Apple Silicon Macs.
- [Set Up on Linux](/docs/linux-setup) covers workstations and servers, including Spark.
- [Set Up on Windows](/docs/windows-setup) covers PowerShell installation.

## Find your way around

| Surface | Use |
| --- | --- |
| [Computers](/console/computers) | Install computers, inspect their state, and manage clusters. |
| [Chat](/console/chat) | Send messages to a loaded model. |
| [Hermes](/docs/hermes) | Use an agent with files and tools, or connect your terminal installation. |
| [Hermes Desktop](/docs/hermes-desktop) | Connect the native app to an Actual computer's managed workspace. |
| [Models](/models) | Browse models and queue downloads to eligible clusters. |
| [User > Settings](/user/settings) | Manage profile, security, connected accounts, and mining. |
| [User > Keys](/user/keys) | Create and terminate API credentials. |
| [Docs](/docs) | Read guides and API reference material. |

[Organizations](/docs/organizations) let people share access to computers. Sharing a computer or placing devices under one account does not itself create a distributed inference cluster.

## More than one computer

Bare `actual` starts a standalone node. Signing into the same account does not automatically configure distributed execution. See [Clustering](/docs/clustering) to select and verify the computers used by an API request.

## Setup with an agent

Point your agent at [agents.actual.inc](https://agents.actual.inc), [llms.txt](https://actual.inc/llms.txt), or the longer [llms-full.txt](https://actual.inc/llms-full.txt). You complete device authorization in your browser; the agent can then inspect status, download and load a model, and test inference.

Online platform and relay use require an authorized device. Explicit `actual --offline` provides local standalone operation without relay, telemetry, or LAN clustering; it can start idle, but inference still needs an installed model.

## If setup stops short

An installed CLI is not proof that a model is serving. Check `actual whoami`, `actual status`, and `actual models list` in that order. An empty inventory needs a download; an installed model needs loading; an offline device needs its daemon and connection restored. See [Logs & Telemetry](/docs/logs-and-telemetry).