Docs › Requirements

Requirements

What Sparehand needs on the Mac you work on (the hub), on each spare machine (a worker), and what the beta features ask for on top. The core needs little: a Mac and one other computer that accepts SSH.

The Mac (the hub)

macOS 14 (Sonoma) or newer, on Apple silicon or Intel: the app is a universal build, with Node.js bundled for both. It is the one machine that needs the app and the licence.
Apple's command-line developer tools, for git. Setup offers to install them (macOS downloads them in a few minutes); nothing else to install.
Your projects as git checkouts. A job copies the checkout to the worker, leaving out what git ignores; folders that aren't git repositories can't be sent.

Spare machines (workers)

Any of these, at home or anywhere with a network path to the hub (a server by IP or domain name, or a machine on your Tailscale network):

	What it needs	How it is set up
A Mac	Remote Login turned on (System Settings → General → Sharing); Apple silicon or Intel. Node, uv and the other tools your projects use.	Add a machine → Pair with a code, one command in Terminal.
A Linux machine or server	An SSH server, `rsync`, `git`, `bash`. On a server Sparehand uses its own build user.	Pair with a code (one command as an administrator, `sudo` or root), or step by step.
A Windows PC	Windows 10 or 11 with WSL2 (Ubuntu by default: a free Windows feature). The pairing command runs in PowerShell as administrator: it turns on WSL if needed, sets mirrored networking, a task that keeps WSL alive for the hub and a firewall rule for its SSH.	Pair with a code.

For every worker: the account's login shell must be sh-compatible (bash, zsh or sh, not fish); an address the hub can reach (same network, Tailscale, or a public address with SSH open to the hub); enough free disk for a copy of each project it runs (jobs stop going to a machine under 5 GB free); and at least 30% of its memory free when a job arrives, or it is skipped.

Graphics cards. sparehand run --gpu, the local picture and clip models and the GPU-side of Studio work on NVIDIA cards only (CUDA), on Linux or on Windows through WSL2 (the NVIDIA driver on the Windows side; WSL's own CUDA libraries). A Mac with Apple silicon runs AI chat models with Ollama on its unified memory, but not the picture or clip models, and never counts as a GPU worker. AMD and Intel graphics aren't used.

Waking machines. Wake-on-LAN needs a wired connection that stays powered in sleep, or a Wi-Fi card that supports waking, and the machine's own setting turned on (Waking and sleeping).

The beta features

Each of these works and is part of the licence, but is newer than the core, leans on outside services or accounts you bring, and may change between versions.

Local AI and cloud AI providers

Local models (Local AI): a worker with an NVIDIA card, or a Mac with Apple silicon; the model's size decides what fits (about 8 GB of graphics memory runs models up to roughly 8 billion parameters). Ollama is installed for that user, no administrator password.
Cloud providers (beta): an account and an API key at each provider you want (Groq, Cerebras, Google Gemini, OpenRouter, NVIDIA, SambaNova, Cloudflare, GitHub, Cohere, Hugging Face; Mistral and xAI are paid). The free tiers have limits the hub counts; some providers refuse some countries.

Studio

A machine with an NVIDIA graphics card for pictures and clips made on your own machines: 8 GB of graphics memory runs the 4B picture model and the 5B clip model; the 14B clip model and the larger picture models want 16 GB or more, and lip sync 40 GB or more. A cloud machine (below) does the same for cents an hour.
About 30 GB of free disk on each Studio machine: ComfyUI with its own PyTorch and the models (--with pictures alone is about 12 GB; clips14b adds about 26 GB). Plus a few GB for Playwright's Chromium, Whisper and a private Python, which every Studio machine gets, the hub included.
API keys for the services a production uses: a Gemini key for writing, voices, pictures, music and clips in the cloud (or Crew with a Claude agent for the writing), and, if you pick them, ElevenLabs, Pexels, Pixabay, Freesound and Higgsfield. A production set to Local only needs none of them, only your machines. Publishing needs a YouTube channel and a Google Cloud OAuth app.
Time: a first picture takes 15 to 25 seconds on an 8 GB card, a 5-second clip a few minutes.

Crew

Accounts for the agents you add: a Claude account (the Master runs on it, Opus or Fable), and any of ChatGPT (through the Codex CLI), Gemini (the Gemini CLI), Kimi or other API keys, Claude Code in the cloud (a Claude account with cloud sessions and a GitHub remote), or a local model on your own machines. Each tool is signed in on the hub with its own login; Sparehand brings none of them.
A git repository with a remote the parts can branch from, and the project's checks runnable on a worker (the Master runs them after each part).

Members

Tailscale on the hub, with the dashboard published on your tailnet (tailscale serve), and each member's device signed in to Tailscale with the login you invite: Using it from anywhere. Inviting by e-mail from the People tab needs a Tailscale API token; Tailscale's free plan allows a few users.
Members use your machines' models, Studio and Crew on prepaid credits you give them (Credits); what they make runs on your machines and, where you allow it, on your cloud and API accounts.

Machines in the cloud

An account at Runpod, Amazon Web Services, Microsoft Azure or Google Cloud and a credential Sparehand may use: an API key (Runpod), an access key (AWS), a service principal or the az CLI's login (Azure), a service account key or gcloud's login (Google). The provider bills you directly; Sparehand shows the price an hour and the cost so far, and stops the machine after its idle minutes.
GPU quota: AWS, Azure and Google start new accounts at 0 GPUs; ask for quota before the first GPU machine. Google also needs billing enabled (a trial can't have a GPU). Runpod needs a few dollars of credit.
The machine's disk or volume: 80 GB or more for a Studio machine with the clip models.

Sessions

Claude Code on the hub, signed in. Sessions started from the dashboard run as your user; the tab reads Claude Code's own transcripts under ~/.claude.
Notifications on your phone need the dashboard open over Tailscale (an address whose certificate the browser trusts; a phone paired on this network gets none) and the browser's push service.

Work items

A Jira Cloud or Azure DevOps account and a personal access token, kept in the hub's Keychain (Work items).