How it works

The inventory, scope, harness and source tags, and what happens while a machine is offline — the concepts behind the dashboard.

Inventory

A machine’s inventory is the list of skills and MCP servers installed on it. Each machine’s agent scans its own filesystem, watches it for changes, and keeps this dashboard’s copy of that list in sync — so Inventory for a given machine reflects what is actually on disk there, not a plan for what should be.

Scope: global vs. project

global means installed for every project on that machine — the ordinary case for most skills and MCP servers.project means installed for one project directory specifically; a project only gets project-scoped scanning once it has been registered with loadout watch <path>.

Harnesses

An item is not owned by a single coding tool. harnesses lists every present client that currently holds a copy of the skill or an MCP entry for that server. Skills live in .agents/skills so any Agent Skills client can read them; adapters also project into native skill directories when those harnesses are present. MCP is written only into configs of harnesses the agent has detected on that machine — Claude Code, Codex, Cursor, Gemini CLI, and GitHub Copilot.

Source: manual, git, npm, marketplace

manual

Found on disk with no recorded provenance — the agent’s scanners cannot know where a manually placed file came from, so anything they discover is tagged this way. Not restorable from a snapshot, since there is no source to re-install it from.

git / npm

Installed by the agent from a git repository or an npm package, with the source reference recorded at install time.

marketplace

Installed from a loadout marketplace listing. Only listings with an approved review status are installable.

The realtime connection

Each running agent subscribes to a private realtime channel scoped to its own machine — no other machine’s channel, and no anonymous client, can read or send on it. A command arriving on that channel is validated against a shared schema before the agent acts on it; anything that fails validation is dropped rather than applied.

Worth knowing

Offline machines catch up

With loadout-agent 0.2.0 or later, a command for an offline machine waits in its queue and runs when the daemon reconnects; the dashboard shows it as queued, then waiting for approval, then done or denied. Older agents only receive live commands, so for them the dashboard disables install, toggle, remove and restore until the machine is online again (loadout service install, or loadout run).

A toggle waits for the machine

Flipping a switch in Inventory sends a command; the switch only moves once that machine confirms it did the work, rather than assuming success right away.

Only approved marketplace listings are installable

The marketplace shows submissions that passed review, so installing from it is picking off a vetted shelf rather than fetching an arbitrary URL.