The short list: Node.js 22+, pnpm 8+, a macOS or Linux machine (Windows works via WSL2), Chrome 121 or newer if you want the browser extension, and one LLM API key (Cerebras' free tier is the recommended default). Everything else the installer detects and offers to set up.
The requirements table
| Component | Requirement |
|---|---|
| Node.js | 22+ |
| pnpm | 8+ (corepack enable pnpm; corepack ships with Node) |
| macOS | Native, Intel and Apple Silicon |
| Linux | Native; the primary development platform |
| Windows | Not native: run inside WSL2 (wsl --install -d Ubuntu); the install refuses up front on native Windows |
| Chrome extension | Chrome 121+ (the CSS Custom Highlight API reached stable in 121; older Chrome renders no highlights) |
| LLM access | One provider key; a keyless free-pool route exists for blanks only |
The native dependency, explained
The runtime sandbox for third-party JavaScript blanks uses isolated-vm, a native C++ binding. Prebuilt binaries cover Linux and macOS on x64 and arm64 plus Windows x64, so on normal machines it installs with no toolchain at all. On rarer architectures the install falls back to compiling, which needs build-essential and python3 on Linux or xcode-select --install on macOS. The installer probes the binding and prints platform-specific guidance if it can't load.
What the installer handles for you
Host-specific tooling (bun for OpenCode and the shell app, tmux for the shell app, plus Linux extras like brightness and speech utilities) is detected at install time with a yes/no/details prompt; contained tools land under ~/.opencues/vendor/ so uninstalling can clean them. Each editor integration pins the version it patches (currently Claude Code 2.1.206, OpenCode 1.14.17, Gemini CLI 0.41.2), and the installer manages those forks itself.
What's deliberately not on the list
No GPU (unless you opt into local inference via Ollama, where the validated default model wants about 1.4 GB of VRAM), no account, no server to run. The repo publishes no minimum disk or RAM figure, so we won't invent one here; in practice the footprint is a Node monorepo plus per-host fork directories. First installs take a few minutes per host; re-runs about a minute.
Related: How to install · LLM providers · Where is it installed? · GitHub repo