System requirements
The operating systems, packages and tools AI Workbench and its nodes need.
AI Workbench runs on Windows, macOS and Linux. The first-run wizard checks the machine and offers to install what is missing, so you rarely need to prepare anything.
The desktop app
| Operating system | Package |
|---|---|
| Windows (x64) | Installer, per user: no administrator rights needed. Signed by Mojo Up. |
| macOS | .dmg (and .zip). |
| Linux (x64) | AppImage and .deb. |
The .deb package depends on libgtk-3-0, libnotify4, libnss3, libxss1, libxtst6, xdg-utils, libatspi2.0-0, libuuid1 and libsecret-1-0, which apt installs with it.
Tools the workbench uses
| Tool | Needed | Why |
|---|---|---|
| Git | Yes | Every repository is a git repository; each agent works in its own branch and worktree. |
| Node.js 20 or later, and npm | Yes, for most agent tools | The agent tools' adapters run on Node.js, and npm installs the agent tools. |
| At least one agent tool | Yes | Claude Code, GitHub Copilot CLI, Codex, Google Antigravity, OpenCode, Kilo Code, Grok Build or Cline, installed and signed in. Or the built-in native bridge with a local or hosted model. |
| Git Credential Manager | No | Signs you in for HTTPS clones and pushes; SSH keys work too. |
GitHub CLI (gh) | No | Signs Copilot in and reads its credits; also opens pull requests when you are not signed in to GitHub in the app. |
Python 3 and uv | No | The full memory engine. A lighter indexer works without it. |
| WSL and a Linux distro | Windows, optional | For repositories that live inside Linux. Installing WSL needs a restart of Windows. |
| Docker (or Podman) | Optional | Only for repositories in a dev container. |
| Homebrew | macOS, optional | Used to install tools. |
The wizard's This computer step and Settings → Setup check these and offer an install button for each.
Local models
A local model is optional. If you want one, the workbench can download its bundled llama.cpp runner (CPU, CUDA or ROCm builds), or use Ollama, LM Studio, Foundry Local, or MLX on Apple silicon. Using a Windows-side local model from a WSL project needs WSL's mirrored networking; the workbench offers to configure it.
Nodes
mojoup-node runs on Linux (x64 and arm64), macOS (Intel and Apple silicon) and 64-bit Windows (x64). It needs git, at least one agent tool signed in as the node's user, and free disk space for agent worktrees: mojoup-node doctor warns below 5 GB. See Install a node.
Sandboxed agent sessions need a Linux node with Landlock (ABI 3 or later), seccomp user notification, Docker and NVIDIA OpenShell. See Sandboxed nodes.
Network
The workbench needs outbound HTTPS to the agent tools' and model providers' own services, and, when signed in, to Mojo Up AI Cloud (https://ai.mojoup.com.au) or your organisation's platform. Updates come from the Mojo Up portal. No inbound port is needed unless you make the workbench a team hub or turn on remote control on your network.