A Claude Skill is a folder of instructions, and optionally scripts and reference files, that Claude loads when a task matches its description. An MCP server is a running service that lets Claude read or change a system it cannot otherwise reach. Use a Skill when Claude has the access but does the task inconsistently. Use MCP when it lacks the access. Many workflows need both.
What is the difference between a Claude Skill and MCP?
A Skill is knowledge. An MCP server is access.
A Skill is a folder with a SKILL.md file: a short YAML header with a name and a description, then instructions. It can also carry scripts, templates and reference documents. Claude reads the full file when a request matches its description. Nothing to host, no sign-in to manage.
An MCP server is a program that speaks the Model Context Protocol. It exposes tools, such as "search contacts" or "create a booking", that Claude can call against a live system, under a real identity. It has to be deployed, secured and kept running like any service.
| Skill | MCP server | |
|---|---|---|
| What it adds | Instructions, conventions, scripts, templates | Tools and data from an outside system |
| What it is | Files in a folder | A running service at a URL, or a local process |
| Sign-in and permissions | None of its own | OAuth or local credentials, enforced by the server |
| Side effects outside the chat | Only through tools Claude already has | Yes, whatever its tools allow |
| Who can write one | Anyone who can describe the task well | An engineer |
| Cost to change | Edit a text file | Change code and redeploy |
How do you decide which one you need?
Ask what Claude got wrong last time.
- It had the information but did it differently each time. Wrong format, skipped a step, ignored your house style. That is a Skill.
- It asked you to paste something in, or guessed. The data lives in a system Claude cannot see. That is an MCP server, or a connector a vendor already publishes.
- It could see the system but used it badly. Called the wrong tool, used the wrong filter, forgot to check a second record. That is a Skill that teaches the procedure for the tools you already have.
- It needs to act somewhere with a named person's permissions. That is MCP, because only a server can enforce who may do what.
Check the connector directory before building anything. If a vendor already publishes a connector, what is left is usually a Skill.
How do Skills and MCP work together?
The Skill is the playbook and the MCP server is the equipment. A good Skill names the tools it expects, the order to call them in and the checks between calls. Anthropic's guidance says to name MCP tools in full, as ServerName:tool_name. A Skill that says "look up the account in the CRM" with no CRM connection gives Claude an instruction it cannot follow.
spARK Labs uses both. A remote MCP server gives Claude access to BossHub, Verkada, Nexudus and Outlook (case study), and the rules for matching founders to advisors live in a Claude skill that proposes pairs for the team to review (case study).
A plugin bundles skills, connectors and sub-agents so a team installs one thing. Claude Cowork, which Anthropic began merging into ordinary Claude conversations in September 2026, runs multi-step tasks using whichever Skills and connectors are installed.
Where do teams go wrong?
- Building a server for what is really a style guide. If Claude already has the data, a server adds hosting and secrets for no gain.
- Hiding access inside a Skill. A script with a pasted API key has no per-user permissions or audit trail. Shared systems belong behind a server with sign-in.
- A description that never triggers. Claude chooses a Skill from its description alone. "Helps with reports" rarely fires. See how to write a SKILL.md.
- One giant tool. An MCP tool that "updates any record" is hard to instruct and to secure. See MCP server design.
Common questions
Do Skills work outside Claude?
Largely, yes. Anthropic published Agent Skills as an open standard in December 2025, and tools including ChatGPT and Codex, Gemini CLI, Cursor, VS Code and GitHub Copilot now read the SKILL.md format. Instructions travel well. Scripts depend on what each tool can run, so test a Skill in every tool you plan to use it with.
Which Claude plans include Skills and custom connectors?
As of September 2026, Anthropic's help center lists Skills on Free, Pro, Max, Team and Enterprise plans with code execution turned on, and custom MCP connectors on the same plans, with Free limited to one custom connector. On Team and Enterprise, an Owner adds connectors and can provision Skills for the whole organization.