The Quick Answer
Use Skills to teach Claude how to do something. Use MCP servers to give Claude access to external systems. Use Plugins when you want a ready-made bundle of skills, agents, hooks and MCP servers installed as one unit.
That's the whole relationship. A plugin is not a competitor to a skill any more than a toolbox is a competitor to a hammer. The confusion comes from timing: MCP was released in November 2024, Agent Skills arrived in October 2025, and Claude Code plugins entered public beta in October 2025, with plugins reaching Cowork in January 2026. Many people installed all three at once and lost track of what each one was for.
What Skills Actually Do
A Skill is a folder of instructions that teaches Claude how to do a specific task the same way every time. It lives in .claude/skills/ (project) or ~/.claude/skills/ (personal) and consists of a SKILL.md file plus optional scripts and templates.
The key detail: Skills load on demand. At startup Claude sees only the name and description of each skill (roughly 100 tokens per skill). When a task matches a description, Claude loads the full instructions. If nothing matches, nothing else loads and your context stays lean. A system prompt or CLAUDE.md, by contrast, sits in context the whole session.
Real use cases:
- Apply brand style guidelines to documents consistently
- Structure meeting notes in company-specific formats
- Generate emails following your template patterns
- Run data analysis workflows your team standardized on
- Create tasks in Linear/Asana/Jira following team conventions
Skills follow the open Agent Skills format, which other agent tools have also adopted. A skill written for Claude Code can usually be reused in another tool that supports the format with little or no change; check that tool's docs for where it looks for skills.
What MCP Servers Do
The Model Context Protocol (MCP) is an open standard for connecting AI tools to data sources and services. An MCP server connects Claude to something outside the conversation: a database, an API, GitHub, Linear, Slack, your design tool, a browser.
When you connect an MCP server, Claude can:
- Fetch data from APIs or databases
- Act in services like GitHub (read PRs, list issues), Slack, or Linear (create tasks)
- Drive a real browser or other running processes
Context cost in 2026: Claude Code now runs MCP tool search by default. At session start only tool names and server instructions load; full tool definitions load when Claude actually needs them. Servers no longer flood your context just by being installed. Full definitions still load upfront if you set ENABLE_TOOL_SEARCH=false, route Claude Code through a non-first-party ANTHROPIC_BASE_URL proxy, or mark a server alwaysLoad: true. Tool results still cost tokens: Claude Code warns when one MCP tool output passes 10,000 tokens and caps it at 25,000 by default (MAX_MCP_OUTPUT_TOKENS).
Local vs. remote: MCP servers run locally (Claude Code starts them as child processes over stdio) or remotely over HTTP. SSE transport is deprecated; use HTTP where the server supports it. Where the connection starts from matters: Claude Code connects from your machine, so a server on your VPN works. A connector added in claude.ai is reached from Anthropic's cloud, so it must be reachable over the public internet.
What Plugins Are
A Plugin is a container. Per the Claude Code docs, a plugin is a directory of skills, agents (subagents), hooks and MCP servers that Claude Code installs and loads as one unit, described by a manifest at .claude-plugin/plugin.json.
This removes the friction of hand-assembling a workflow. When Anthropic brought plugins to Cowork in January 2026, it open-sourced 11 plugins built by its own teams for functions such as sales, finance, legal and data. A sales plugin, for example, might bundle a prospecting skill, research commands, a CRM connector and a subagent for competitive analysis. A new hire installs the plugin and Claude already knows how the team works.
Plugins are not new capabilities. Everything inside a plugin could be set up by hand; a plugin saves the assembly and gives you versioned updates from its marketplace. Claude Code adds Anthropic's official marketplace (claude-plugins-official) the first time you start an interactive session: run /plugin and browse the Discover tab. Plugins in the official marketplace show a context-cost estimate before you install.
Keep in mind that an enabled plugin is part of every session: the names and descriptions of its skills and agents sit in context on every turn, and its MCP servers and hooks run alongside each session. Disable plugins you are not using.
Head-to-Head Comparison
| Dimension | Skills | MCP Servers | Plugins |
|---|---|---|---|
| What it does | Teaches Claude how to do something | Connects Claude to external systems | Bundles skills, agents, hooks and MCP servers |
| Context overhead | ~100 tokens per skill (name + description); full instructions load only when relevant | With tool search (default): tool names and server instructions at start, full definitions on demand; tool results count when used | Sum of what it bundles; names and descriptions of its skills and agents load every turn |
| Setup | Write a SKILL.md | claude mcp add or a .mcp.json file | /plugin → Discover, or /plugin install name@marketplace |
| What you provide | Procedural knowledge: steps, conventions, patterns | Access to external data, APIs, tools | A packaged, versioned setup |
| Works beyond Claude Code? | Yes, in tools that support the Agent Skills format | Yes, MCP is an open standard with many clients | Same format installs on claude.ai and Cowork, with different component support |
| Shared across team | Commit .claude/skills/ or ship in a plugin | Commit .mcp.json (project scope) | Project-scope install via committed .claude/settings.json |
| Use if you need to… | Standardize workflow, capture know-how, teach Claude your style | Integrate external systems, APIs or services | Install or distribute a complete domain workflow |
Decision Guide: When to Use Each
Use Skills when:
- You want Claude to follow your team's workflow or conventions (code review style, document format, brand guidelines)
- You're encoding procedural knowledge that should be consistent across team members
- You want to save context by loading instructions only when they're relevant
- You need to share patterns across multiple AI tools (Claude Code, Cursor, Copilot, etc.)
Use MCP servers when:
- Claude needs to read or write to files, databases, or APIs
- You need Claude to interact with external services (GitHub, Linear, Slack, Notion, etc.)
- You're building integrations that require real-time data access
- You want to give Claude "hands" to do work outside the conversation window
Use Plugins when:
- You need a complete, pre-configured workflow for a specific domain (sales, legal, finance, engineering)
- Your team is onboarding and you want to ensure everyone has the same setup
- You want to avoid hand-assembling skills, MCP configs, and commands
- You trust the plugin author and have verified the included components
A realistic example: Your engineering team standardizes on code review. You write a Skill that teaches Claude your team's review checklist and style. You connect a GitHub MCP server so Claude can read PRs and leave comments. You bundle both into a Plugin called "code-review" that the whole team installs. New hire runs /plugin install code-review@your-team-marketplace and they're ready to go.
Can You Combine Them?
Yes, and you should. They work together: a Skill can tell Claude when and how to use MCP tools, and a Plugin can bundle both.
Realistic setup:
- 2–3 MCP servers (GitHub for code, a docs server such as Context7, one domain-specific service)
- 3–5 custom Skills (your team's workflow, review conventions, document format, analysis patterns)
- 1–2 Plugins for your role, from a marketplace you trust
Skills teach Claude how to work. MCP servers give it reach. Plugins remove the assembly work. Most teams end up using all three, each for what it is actually good at.
Warning: a plugin runs with your user privileges, whatever marketplace it comes from. Anthropic's docs separate official, community and third-party marketplaces; anything outside Anthropic's own is third-party. Read the plugin's source, its hooks and its MCP servers before installing. The plugin security guide explains what to review.