- 05
- June
When a team starts using Claude Code and wants to extend it, four similar-sounding terms cause confusion — Claude Code, Plugins, MCP and Skills. Many assume you must pick one, but all four sit at different layers and work together. This article lays out a clear mental model: Claude Code is the host, MCP is the "pipe," Skills are the "instructions," and Plugins are the "package" that bundles everything.
The mental model, in short
- Claude Code = the coding agent (CLI/IDE) — the host that runs everything
- MCP = transport — connects Claude to external systems (DB, API, GitHub) as callable tools
- Skills = instructions — a SKILL.md that teaches Claude "how to do this task," loaded on demand
- Plugins = package — bundles skills + MCP + hooks + subagents into one installable/shareable unit
What each one is
| Component | Role | Analogy |
|---|---|---|
| Claude Code | Coding agent on CLI/IDE — reads/edits files, runs commands | the vehicle (host) |
| MCP | Protocol connecting to external systems, exposing tools via a standard interface | the pipe (transport) |
| Skills | SKILL.md teaching workflows/knowledge, loaded when needed (tiny token cost until invoked) | the manual (instructions) |
| Plugins | A package bundling skills + hooks + subagents + MCP servers, installable/shareable via marketplace | the product box |
Deep dives: Claude MCP, Claude Skills, and Managed Agents / Subagents.
The mental model — how they work together
The easiest way to remember: three layers with different jobs, and Plugins is what wraps them.
- MCP = transport — connects to external systems (e.g. an ERP database)
- Skills = instructions — tells Claude "what to do with it" (e.g. a month-end close procedure)
- Plugins = product — bundles MCP + Skills + more into one unit to share/reuse
They compose — e.g. a Skill calls an MCP tool, and both are packaged inside one Plugin.
When to use which
| Situation | Use |
|---|---|
| Customize Claude for your own project/team | Skill (SKILL.md in .claude/skills/) |
| Connect to external systems (DB/API), shared across agents, with auth/rate-limiting | MCP server |
| Reuse the same setup across repos / distribute to others | Plugin |
| Just one repo, small team | Keep it in .claude/ |
Enterprise use-case examples
- MCP: build an MCP server connecting Claude to your ERP's PostgreSQL to query/summarize data securely (with central auth).
- Skill: write a SKILL.md "account reconciliation procedure" so anyone can invoke
/reconcile. - Plugin: bundle MCP (ERP) + Skills (accounting/inventory workflows) + hooks into one plugin and distribute it so every implementation team uses the same setup.
- Subagent: use it to split big tasks into sub-tasks (see Managed Agents).
Summary
You don't pick one — all four sit at different layers. Start with a Skill (fastest) for team tasks, add MCP when you need system connectivity, then package as a Plugin when you want org-wide reuse.
The right question isn't "Claude Code, MCP, Skill or Plugin — which is better," because they're at different layers. MCP is the pipe, Skills are the instructions, Plugins are the box that wraps them. Once you grasp what each does, designing enterprise AI workflows becomes systematic.
- A technical view on extending Claude in the enterprise
References
- Claude Code Docs — Extend Claude Code (features overview)
- Model Context Protocol (MCP) — Official site
- Claude Code Docs — Agent Skills
Want to connect Claude to your ERP / internal systems?
Grand Linux has over 24 years of experience integrating Thai enterprise systems — we help design MCP/Skill/Plugin setups to connect Claude to your ERP and internal systems securely (optional paid service).
See Claude IntegrationTel 02-347-7730 | sale@grandlinux.com
