Discover, install and manage the skills, agents and MCP servers your AI tools use — from a desktop app instead of scattered folders and hand-edited JSON.
Runs on macOS, Windows and Linux.
Claude Code already installs plugins, but through a terminal list, and only for content packaged as a formal plugin. Everything else — a skill you dropped in a folder, an MCP server you configured months ago — is invisible: no version, no origin, no way to tell whether it still works.
Loadout adds the part that was missing:
- An inventory of what is actually on your machine, whether or not Loadout put it there
- A catalog you can browse, with the README rendered on screen instead of a name in a list
- A life cycle: update, revert, disable, and a clear answer to "did this change upstream, or did I edit it?"
Scans skills, agents, commands and plugins, plus MCP servers both global and per project. Flags what needs attention: dangling links, folders that disappeared, a required variable with no value set.
Reads the Claude Code marketplace.json format, so it works with the catalogs
that already exist, and a richer format of its own. Several sources at once,
added by URL or imported from a file, checked before being added and cached so
the app still works offline.
Nothing is written before you see what is coming: the confirmation lists every file, the destination, and the capabilities declared in the content — whether it runs commands, writes files or reaches the web.
Content is cached by commit, so the same version is never downloaded twice, and installed as a link, so its origin stays traceable.
A configuration form generated from the catalog, and — the part that saves the most time — a connection test before saving. Loadout starts the server, performs the MCP handshake and lists the tools it exposes. The usual cycle is edit JSON, restart the client, and only then find out it was wrong.
Grab the installer for your system from the releases page.
macOS: this build is not signed yet, so the first launch needs System Settings → Privacy & Security → Open anyway, and the app cannot install its own updates — it points you to the download page instead. Windows and Linux update normally.
npm install
npm run dev| Command | What it does |
|---|---|
npm run dev |
app with hot reload |
npm run typecheck |
TypeScript across main, preload and renderer |
npm test |
test suite |
npm run build |
typecheck plus bundle |
npm run dist |
installers for the current system |
Releasing is driven by main: raise the version in package.json, add the
matching notes under changelog/, and push. The workflow builds on the three
systems, publishes to GitHub Releases and creates the tag. Pushing again with
the same version does nothing, so only a version bump triggers a release.
A catalog is a JSON file served over HTTP. Minimal shape:
{
"schemaVersion": 1,
"name": "My catalog",
"items": [
{
"id": "owner/repo#my-skill",
"name": "my-skill",
"type": "skill",
"description": "What it does, in one sentence.",
"source": {
"kind": "git-subdir",
"url": "https://github.com/owner/repo",
"ref": "main",
"subpath": "skills/my-skill"
}
}
]
}description and readme also accept a map by language
({ "en": "...", "pt-BR": "..." }). Items of type mcp can declare runtime
and configSchema, which is what generates the configuration form.
See catalogs/loadout.json for a complete example.
The app installs third-party code on your machine, so a few things are deliberate:
- Markdown from catalogs is sanitized, and remote images are not loaded — opening an item never pings someone else's server
- Values of MCP environment variables and headers never leave the main process; the interface only ever receives the key names
- Downloaded archives are filtered before extraction, so no entry can write outside its destination
- MCP configuration goes through the official Claude Code CLI, never by editing its files directly, and the file is backed up first
Content comes from anthropics/skills, claude-plugins-official and modelcontextprotocol/servers, each under its own license.
Loadout is an independent project and is not affiliated with Anthropic.
MIT © Kaio Piola