Claude Skills vs MCP Connectors vs Plugins: What Is the Difference and Which Should You Use?

Originally published at devtoolpicks.com

If you have spent any time with Claude recently, you have probably run into these three terms and wondered what makes them different. Skills, MCP Connectors, Plugins. They sound related, they overlap in confusing ways, and Anthropic has not published a single clear diagram showing how they fit together.

This post is that diagram, in words. By the end you will know exactly what each one does, when to use it, and which one you actually need right now.

Quick Verdict

The Simple Mental Model

Before going deeper, here is the framing that makes everything click:

MCP is the plumbing. It connects Claude to the outside world.

Skills are the instructions. They teach Claude how to do specific things.

Plugins are the finished product. They bundle both together into something you install once.

They are not competing options. They are three layers of the same stack. A Plugin can contain Skills. A Skill can use MCP tools. Understanding which problem you are solving tells you which layer you need.

View the interactive diagram on devtoolpicks.com

Read it left to right: a Plugin bundles Skills, MCP Connectors, and Slash Commands together. Skills and MCP both feed Claude. MCP is the only layer that reaches out to external services.

Claude Skills

A Skill is a plain Markdown file (called SKILL.md) that gives Claude specific instructions, context, or a repeatable workflow for a particular task.

Think of it as onboarding documentation for Claude. Instead of re-explaining your preferences, your code conventions, or your review process every single chat, you write it once in a Skill and Claude loads it automatically when relevant.

What Skills are good at:

• Teaching Claude your team's coding conventions
• Defining a repeatable workflow (deploy process, PR review checklist, blog post format)
• Giving Claude domain knowledge it would not have by default
• Keeping your main context window clean (Skills only load when needed, costing roughly 30-50 tokens each)

A real example: You review pull requests every week and always want Claude to check the same things: naming conventions, missing tests, security issues. Instead of typing that checklist into every new chat, you create a Skill called pr-reviewer with those instructions. Claude picks it up automatically whenever you ask it to review code.

The honest con: Skills are instructions only. They cannot fetch data, call APIs, or connect to external services. If you need Claude to actually read your GitHub issues or check your database, a Skill cannot do that on its own. That is what MCP is for.

Who should NOT use Skills: If your problem is connecting Claude to an external service, Skills are the wrong tool. They shape behavior, not connectivity.

MCP Connectors

MCP stands for Model Context Protocol. It is an open standard (now under the Linux Foundation, co-founded with Block and OpenAI) that defines how AI models connect to external tools and data sources.

The USB-C analogy keeps coming up for good reason: build an MCP server once and it works with Claude, ChatGPT, Gemini, or any other tool that speaks the protocol. OpenAI adopted it in March 2025. Google DeepMind followed in April.

In the Claude context, Connectors are the pre-built MCP integrations Anthropic and third parties have packaged for common services: Microsoft 365, Google Drive, Slack, GitHub, Linear, and more. When you enable a Connector, Claude can actually read your emails, access your files, and interact with those services on your behalf.

What Connectors are good at:

• Giving Claude real-time access to external data (emails, documents, repositories)
• Acting as you with delegated permissions via OAuth (if you cannot see a SharePoint doc, Claude cannot either)
• Connecting Claude to your actual workflow without writing code

The honest con: MCP Connectors consume a significant amount of context before you type a single word. A five-server MCP setup can use around 55,000 tokens of context overhead upfront. That is a third of your context window gone before the conversation starts. Install only what you actually use.

Who should NOT use Connectors: If you just want to teach Claude how to follow your workflow or coding style, you do not need MCP. That is a Skill.

Plugins

Plugins are the newest addition, shipping with Claude Cowork on January 30, 2026. They landed hard enough to wipe $285 billion off software stocks on the day of announcement.

A Plugin bundles Skills, MCP Connectors, slash commands, and sub-agents into a single installable package. If Skills are individual recipe cards and MCP is the kitchen plumbing, a Plugin is the full kitchen, stocked and ready to cook.

Anthropic launched 11 official Plugins at release, covering legal, sales, finance, marketing, data analysis, and more. A Plugin marketplace launched in February 2026.

What Plugins are good at:

• Distributing a complete workflow to a team without requiring anyone to configure things manually
• Packaging a job-specific toolkit (a legal Plugin includes NDA review workflows, document management connectors, and slash commands like /review-contract)
• Ensuring everyone on a team has the same setup with one install

A real example: A sales team installs a Plugin that connects to their CRM via MCP, includes Skills for their prospecting workflow, and exposes commands for research and follow-up drafts. New hire joins, installs the Plugin, and Claude already knows how the team works.

The honest con: Plugins are overkill for individual use. If you are a solo developer or indie hacker, you do not need to package and distribute anything. Just write Skills directly for your own workflows and add MCP servers for external services you actually use.

Who should NOT use Plugins: Solo developers working alone. The overhead of building a Plugin is only worth it when you need to distribute a setup to others.

How They Fit Together in Practice

These three layers are not mutually exclusive. Real setups combine them:

A developer building a SaaS might have:

• A Skill that teaches Claude their codebase conventions and PR review checklist
• An MCP Connector for GitHub so Claude can read actual issues and open PRs
• No Plugin needed, because they are working alone

A small team might have:

• A Plugin that packages their deployment workflow Skill, a GitHub MCP Connector, and a /deploy slash command
• One install and the whole team is set up identically

The decision is simpler than it looks. Ask yourself three questions:

1. Do I need Claude to connect to an external service? Yes: use an MCP Connector. No: skip it.
2. Do I have a repeatable workflow or specific instructions I keep re-explaining? Yes: write a Skill. No: skip it.
3. Do I need to distribute this setup to others? Yes: package it as a Plugin. No: skip it.

Which One Do You Actually Need Right Now?

You are a solo developer or indie hacker using Claude for coding:
Start with Skills. Write a SKILL.md with your project conventions, preferred patterns, and any workflows you repeat. Add a GitHub MCP Connector if you want Claude to actually interact with your repository. You do not need a Plugin.

You are using Claude for content or writing:
Skills are your best tool. Define your voice, your post structure, your review checklist. Claude will apply it consistently without you repeating yourself.

You want Claude to read your emails, files, or documents:
MCP Connectors are what you need. Enable the Google Drive or Microsoft 365 connector and Claude can access your actual files.

You are setting up Claude for a team:
Build a Plugin. Package your Skills, Connectors, and any slash commands into one install so everyone gets the same setup without manual configuration.

FAQ

Do Skills work in both Claude.ai and Claude Code?

Yes. The same SKILL.md format works across Claude.ai, Claude Code, and the API. Skills have been adopted as an open standard and also work with Codex and Gemini CLI.

Do I need to code to use MCP Connectors?

No. Pre-built Connectors for common services like Google Drive, Microsoft 365, and GitHub are available through Claude's settings without writing any code. Custom MCP servers require development work.

Are Plugins free?

The Plugin marketplace launched in February 2026. Anthropic's official Plugins are free and open-source on GitHub. Third-party Plugins vary by provider.

What is the difference between Skills in Claude.ai and CLAUDE.md in Claude Code?

They serve the same purpose but in different contexts. CLAUDE.md is for Claude Code (the terminal tool) and applies to a specific project. Skills in Claude.ai work across your conversations. Both give Claude persistent instructions without requiring you to repeat yourself every session.

Should I be worried about security with third-party Skills?

Yes, with caution. Snyk's ToxicSkills research from February 2026 found that 13.4% of publicly available Skills had critical vulnerabilities. Skills can include executable scripts. Read any SKILL.md file you install from an external source before enabling it, the same way you would review an npm package.

The Bottom Line

The confusion is understandable. These features rolled out at different times, work slightly differently across Claude.ai and Claude Code, and the names do not make the hierarchy obvious.

But the mental model is clean once you have it:

Skills teach Claude how to work. Write them for any workflow you repeat more than twice.

MCP Connectors give Claude access to the outside world. Install them when you need Claude to read your actual data.

Plugins distribute everything as a kit. Build them when you are setting up multiple people.

Most solo developers only need Skills and one or two Connectors. Plugins come later, if at all.

If you are evaluating Claude as part of a broader AI tool decision, the ChatGPT Pro vs Claude Max comparison and the Zapier vs Make vs n8n automation breakdown cover how Claude fits into the wider solo developer stack.