Summary
Claude Code’s plugin system packages an entire local development setup — skills, sub-agents, hooks, MCP server configs, LSP servers, background monitors, and a settings slice — into a single versioned, installable unit that can be distributed to a team or published to a marketplace. This official Anthropic tutorial covers both sides: consuming plugins others have built and authoring your own.
Installation runs through `/plugin install organization/plugin-name` in a session. Teams can register a private marketplace once with `/plugin marketplace add` to enable centralized discovery and automatic updates. The video opens with a direct security warning: a plugin runs with the installing user’s full privileges, and every hook it ships fires on every matching tool call. A community plugin can include a `stop` hook that calls out to an external network endpoint with no warning in the UI. The recommendation is to audit every hook, agent, and MCP server before installing — even plugins that have passed Anthropic’s automated review.
Plugin components don’t overwrite existing configuration; they stack alongside it. Pre-tool-use hooks from the plugin and from the user’s own setup both fire on every tool call. Skills and agents are namespaced under the plugin name to avoid clashes. The `settings.json` slice a plugin can ship is intentionally narrow: only `agent` and `subagentStatusLine` keys are honored. Authoring follows the existing `.claude` directory shape, with a `cloudplugin/plugin.json` manifest holding name (the only required field), version, description, and author.
📺 Source: Claude · Published September 03, 2026
🏷️ Format: Tutorial Demo
