Claude skills vs agents vs MCP servers
There are three ways to extend Claude: skills are instruction files that teach Claude how to do something, agents are specialist workers Claude delegates jobs to, and MCP servers are external programs that give Claude new tools — like databases, browsers, and APIs. They stack together, and all three install in about a minute.
| Skill | Agent | MCP server | |
|---|---|---|---|
| What it is | Instruction file (SKILL.md) | Specialist worker (markdown file) | External program exposing tools |
| Where it lives | ~/.claude/skills/ | ~/.claude/agents/ | Runs as a separate process (stdio or HTTP) |
| How it activates | Auto — when your request matches | Delegated — Claude hands it a job | Called — Claude invokes its tools like an API |
| What it changes | How Claude does things | Who does the work | What Claude can reach |
| Install | One curl command | One curl command | claude mcp add … (may need API keys) |
| Example | WCAG accessibility audit checklist | Code-review subagent | GitHub, Supabase, Playwright |
In plain English
What is a Claude skill?
A Claude skill is a folder containing a SKILL.md file — YAML frontmatter (name, description) plus markdown instructions. Claude Code loads skills from ~/.claude/skills/ on startup and automatically activates the right one when your request matches its description. Skills teach Claude how to do something: a workflow, a coding standard, a domain playbook.
What is a Claude agent (subagent)?
An agent is a markdown file in ~/.claude/agents/ that defines a specialist worker — its own system prompt, tool permissions, and focus. Claude delegates whole jobs to it ("review this PR", "research this topic") and the agent works in its own context, then reports back. Skills change how Claude behaves; agents are separate workers Claude can hand tasks to.
What is an MCP server?
An MCP (Model Context Protocol) server is an external program — run locally over stdio or hosted over HTTP — that exposes new tools Claude can call, like querying a database, controlling a browser, or reading Linear tickets. Skills tell Claude what to do; MCP servers give Claude new things it can do.
Should I use a skill, an agent, or an MCP server?
Use a skill when Claude already has the abilities and just needs the know-how (e.g. a code-review checklist). Use an agent when you want to delegate a multi-step job to a specialist with its own context. Use an MCP server when Claude needs to reach a system it can't touch otherwise (APIs, databases, apps). They stack: a skill can orchestrate tools provided by an MCP server.
What is a skill bundle?
A bundle is AugmentClaude's curated pack of skills that work well together for one job — like web design or data engineering. One paste-able command installs the whole set into ~/.claude/skills/.
Want a head start? Curated bundles pack the best skills for a job into one install command — or read how to install Claude skills.