---
name: set-up-skulk
description: Install and configure Skulk on a supported Mac or Linux machine, bring nodes online, choose a compatible model and verify first inference. Use for initial setup or adding a node; preserve an existing installation and the user's chosen release channel.
---

# Set up Skulk

Help the user reach a working node and one useful model response. Read [installation](references/install.md) for OS-specific procedures and [first use](references/intro.md) for model selection and cluster boundaries. These are public product docs bundled from the exact Skulk revision in `source-manifest.json`; they are a snapshot, not a promise that a packaged download has the same version as next-channel docs. Resolve version drift against https://docs.foxlight.ai/ before installing.

## Choose the installation

Inspect the OS, CPU/GPU architecture, available memory and disk, existing Skulk package/runtime, and intended stable or development channel. Preserve a working installation and its settings. Use packaged applications for ordinary macOS and Ubuntu/Debian setup; a source install is appropriate when requested or when the documented packages do not cover the system. A Linux ARM NVIDIA machine is not an x86_64 NVIDIA machine. Windows installation is not covered by these references.

On Apple Silicon, follow the signed app/Homebrew procedure and start Skulk from its menu. On Ubuntu/Debian, choose the desktop package or runtime-only user service as appropriate. Installation does not itself prove the runtime has started. Use the documented entrypoint for that package; do not install a second competing source runtime to fix a stopped desktop app. Let the user enter administrator passwords and OS permission prompts through their normal interface.

## Reach a ready node

Open the dashboard through the app or the configured API origin. Confirm the expected machine in the topology. One node is a complete cluster. Add further machines only when needed, using matching versions and namespaces. A shared namespace does not grant every model compatibility with every engine or GPU.

Use dashboard Settings for normal Skulk configuration. Keep existing model storage and networking choices. For discovery failures, check startup state, permissions, version and namespace before changing networking. Use app logs, the documented user-service logs or node doctor to diagnose a concrete failure; do not keep reinstalling without new evidence.

## Run the first model

Use Model Store → Find Models and inspect the actual card, memory estimate and placement preview. Choose a model appropriate to the user's task and hardware; a catalog listing, downloaded weights or an available engine package does not mean ready inference. Account for artifact and engine downloads and the user's requested scope before starting them. Use an existing ready model when it already meets the task.

If a gated download needs access, have the user supply their own model-read token through the documented settings. Do not put it in source, logs or copied examples. Observe download and placement until the instance is ready, then send a short real request through dashboard Chat or the documented API first-success flow.

## Verify and hand over

Report the actual installed version/channel, the ready node(s), chosen model and successful response. If a task is blocked, say whether installation, service startup, discovery, download, placement or inference failed, with the relevant diagnostic and next action. Distinguish reading a guide, probing an endpoint and generating useful output. Offer client integration as the next step when the user wants it; setup does not require building a capability.