MCP vs Claude Skills

Skills package know-how; MCP packages connections. A Claude Skill is a folder — a SKILL.md of instructions plus optional scripts and reference files — that Claude loads into context when relevant, teaching it how to do something. An MCP server is a running program that gives it live access to external systems. One teaches, the other connects, and they're at their best together.

Side by side

Claude SkillsMCP
What it isA folder of instructions and assets (SKILL.md + files)A protocol, spoken by running server programs
Gives the modelProcedural knowledge — how to do a task wellLive capabilities — data to read, actions to take
LoadingProgressive disclosure: name and description always visible, full body only when the task calls for itTool schemas from every connected server sit in context
RuntimeNo process of its own — files read into context; bundled scripts run in the agent's sandboxA separate process on your machine, or a service at a URL
AuthNone — it's contentOAuth flows or API keys; the server holds real credentials
DistributionCopy the folder; version it in gitInstall a package or add a URL per client
Works inClaude apps, Claude Code and the Agent SDK — an Anthropic convention others are adoptingAny MCP client from any vendor — the full list

When to use which

Encoding how your team does something? A Skill. Brand guidelines, a review checklist, the exact shape of your weekly report — knowledge problems, solved with markdown, no server to run.

Reaching a system Claude can't touch? MCP. No amount of instruction lets a model query your Postgres or file a Jira ticket; that takes a connection with credentials.

Automating a real workflow end to end? Both, split cleanly: MCP servers provide the verbs, the Skill provides the recipe. "Pull this quarter's numbers, build the deck, post it to Slack" is two or three MCP connections and one Skill that knows the order and the standards.

Common questions

Do Skills replace MCP?

No. A Skill can't hold credentials, maintain a connection or call an external API by itself — it's instructions, not infrastructure. What Skills do replace is a chunk of what people previously mis-used MCP for: stuffing static how-to knowledge into tool descriptions. Procedures belong in Skills; connections belong in MCP.

When should I write a Skill instead of an MCP server?

When the thing you want to add is knowledge: a house style, a report format, a workflow, a checklist. If Claude could do the task given the right instructions and its existing tools, that's a Skill — a folder of markdown, no code to run, no server to host.

Can a Skill use MCP servers?

Yes, and it's the natural pairing: the MCP server contributes the tools (query the data warehouse, post the message) and the Skill contributes the procedure (which tables matter, what the report must look like, in what order to do it).

Do Skills work outside Claude?

The format is deliberately simple — a folder with a SKILL.md — and Anthropic has promoted it as an open convention, with other tools adopting compatible 'agent skills'. But it's model-context tooling, not a wire protocol: there's no cross-vendor interoperability contract the way there is with MCP.

More pairings: the compare hub · protocol basics: What is MCP?