# Claude Skills vs MCP vs Connectors vs Plugins (2026)

*Verified against Anthropic official documentation (June 13, 2026 snapshot). Canonical page: https://tygartmedia.com/claude-skills-vs-mcp-vs-connectors-vs-plugins/.*

> **Canonical URL:** https://tygartmedia.com/claude-skills-vs-mcp-vs-connectors-vs-plugins/
> **Full Guide PDF:** https://tygartmedia.com/wp-content/uploads/2026/10/tm-claude-skills-vs-mcp-vs-connectors-vs-plugins.pdf

---

> **Direct Answer:** **Skills** teach Claude *how* to do a task (procedural knowledge in `SKILL.md`), **MCP servers** and **Connectors** give Claude *access* to external tools and data, **Plugins** bundle any combination of commands, skills, subagents, MCP servers, and hooks for Claude Code, and **Hooks** + **Slash commands** deterministically control the session.

## 1. Decision Matrix

| Type | Core Role | Runs Where | Install / Configure |
| --- | --- | --- | --- |
| **Skill** | Teaches **HOW** (`SKILL.md` + optional scripts/templates) | Claude code execution VM | claude.ai zip (`Settings > Features`), `/v1/skills`, or `~/.claude/skills/` / `.claude/skills/` |
| **MCP Server** | Gives **ACCESS** ("USB-C port for AI" via tools, resources, prompts) | Local `stdio` or remote HTTP (SSE deprecated) | `claude mcp add` (local, project `.mcp.json`, or user `~/.claude.json`) |
| **Connector** | Gives **ACCESS** (remote MCP packaged in Claude apps/Cowork/Desktop) | Reached from Anthropic's cloud infrastructure (public URL required) | `Customize > Connectors` or `+` menu (Free users: 1 custom connector limit) |
| **Plugin** | **Container** bundling commands, subagents, MCP servers, hooks, skills | Claude Code (terminal & VS Code) | `/plugin` and `/plugin marketplace add user-or-org/repo-name` (`plugin-name:skill-name`) |
| **Hook** | **Deterministic lifecycle control** (shell/HTTP/prompt) | Claude Code events (`SessionStart`, `UserPromptSubmit`, `PreToolUse`, `PostToolUse`, `Stop`) | `~/.claude/settings.json`, `.claude/settings.json`, or plugin |
| **Slash Command** | Controls session (`/help`, `/compact`, custom commands merged into Skills) | Claude Code session | `.claude/commands/<name>.md` or `.claude/skills/<name>/SKILL.md` |

## 2. Skills Progressive Disclosure & Constraints

- **Level 1 (Metadata, always loaded):** ~100 tokens per Skill (`name` <= 64 chars, lowercase/digits/hyphens, no `"anthropic"`/`"claude"`; `description` <= 1,024 chars).
- **Level 2 (Instructions, on trigger):** `<5k` tokens (`SKILL.md` body).
- **Level 3+ (Resources, on demand):** Effectively unlimited bundled files read/executed via bash. **Important:** custom Skills do not sync automatically across claude.ai, API, and Claude Code.
