Extension types, explained

Like a USB accessory: plug it in and Claude gains a new capability.

Runs as its own process (locally or remote) and talks to Claude over the Model Context Protocol. Gives Claude tools it doesn't have natively — query your database, drive a browser, call the GitHub API, read Slack. Works in Claude Code, Claude Desktop, and any MCP-compatible client.

Install:

claude mcp add <name> — usually an npx or uvx one-liner from the entry's install box.

Examples:

context7 feeds Claude up-to-date library docs; github-mcp-server lets it manage repos, issues and PRs.

Like a browser-extension store bundle: one install, several features.

The packaging format for Claude Code. A plugin can contain slash commands, subagents, hooks, skills and MCP server configs in one installable unit, distributed through plugin marketplaces (the official Anthropic one or any GitHub repo).

Install:

/plugin marketplace add owner/repo then /plugin install <name> inside Claude Code.

Examples:

claude-hud adds a live status display; anthropics/claude-plugins-official is the curated first-party directory.

Like a recipe card Claude pulls out when the dish comes up.

A folder with a SKILL.md (instructions, optionally scripts and resources) that Claude loads only when relevant — keeping context lean. Skills teach Claude how to do a specific task to a high standard: build slides, review accessibility, write SEO copy. Portable across Claude Code, claude.ai and API agents.

Install:

Often bundled in a plugin, or cloned into ~/.claude/skills/ (project: .claude/skills/).

Examples:

anthropics/skills is the official collection; playwright-skill teaches browser-test automation.

Like a specialist contractor your general builder calls in.

A separate Claude instance with its own system prompt, tool allowlist and context window. The main agent delegates bounded tasks to it (review this code, research this question) and gets back the result — keeping the main context clean and enabling parallelism.

Install:

Markdown definitions in .claude/agents/ (or bundled in plugins).

Examples:

VoltAgent/awesome-claude-code-subagents ships 100+ ready-made specialists.

Like a tripwire: fires every time, no AI judgment involved.

A shell command Claude Code executes automatically at lifecycle events — before/after each tool call, on session start, on stop. Deterministic by design: guaranteed guardrails (block dangerous commands), formatting, logging, notifications. Configured in settings.json.

Install:

JSON entries in ~/.claude/settings.json or via a plugin.

Examples:

claude-code-safety-net blocks destructive git commands before they run.

Like a saved text-snippet with superpowers.

A reusable prompt template you invoke explicitly by typing /name in Claude Code, optionally taking arguments. Good for codified workflows you run on demand — /review, /deploy-checklist, /write-tests.

Install:

Markdown files in .claude/commands/ (or bundled in plugins).

Examples:

Most command packs ship inside plugins — see the Plugins tab.