Guide
Claude Skills vs MCP: which do you need?
The short version: a skill changes how Claude behaves, an MCP server changes what Claude can reach. Skills are folders of instructions loaded when a task matches; MCP servers are running processes that expose external tools and data. If your problem is 'Claude should do this differently,' you want a skill. If it's 'Claude can't touch that system,' you want MCP.
These two get conflated constantly, partly because both extend Claude and both arrived within a year of each other. I use both daily and ship a skill, so here's the distinction as it plays out in practice rather than in documentation.
What is a skill?
A folder with a SKILL.md inside: instructions, optionally scripts and assets. Claude reads every installed skill's short description and pulls the full instructions in only when a conversation matches. Skills are files, not processes; nothing runs until Claude uses them. They follow the open Agent Skills standard, so the same folder increasingly works across other coding agents too. The full anatomy is in the plain-English guide.
What is an MCP server?
A running process that speaks the Model Context Protocol and exposes tools and data to Claude from outside: a database, a calendar, a CRM, a scraping service. Where a skill is knowledge Claude carries, an MCP server is a door Claude can open. It exists because models can't reach into your systems on their own, and shouldn't be able to without something governing the connection.
The decision test
Ask which sentence describes your gap. "Claude technically can do this, but does it wrong, generically, or inconsistently": that's behavior, and behavior is a skill. Review checklists, house styles, capture habits, file formats. "Claude literally cannot see or touch the thing": that's access, and access is MCP. Your ticket queue, your analytics, your inbox.
A concrete pair from my own stack: JotThread is pure skill territory, since capturing a decision into a Markdown file is behavior plus a rendering script, with no external system involved. Publishing a draft to my newsletter platform is MCP territory, since that requires an authenticated connection to a real service.
They stack, and that's the mature setup
The best workflows use MCP for the connection and a skill for the judgment. The MCP server gets Claude into the system; the skill encodes what "done right" means once it's there: which fields matter, what the output format is, what never to touch. Access without encoded judgment produces confident messes; judgment without access produces beautifully formatted advice about systems Claude can't see.
If you're choosing what to install first, start with skills, and start small: the roundup by use case covers what's worth having today.
Common questions
Is MCP replacing skills, or the other way around?
Neither. They solve different layers and combine cleanly: MCP connects Claude to a system, and a skill can encode how your team wants that system used.
Do skills require writing code?
No. The minimum skill is one Markdown file of instructions. Code enters only when part of the job must be exact, like parsing a file or rendering a template, and then a bundled script beats hoping the model improvises consistently.
Which is easier to set up?
Skills, by a wide margin: a folder you drop into place. MCP servers are running processes with configuration and, often, authentication, which is reasonable for connecting real systems and overkill for encoding preferences.
A skill, doing skill-shaped work.
JotThread is behavior in a folder: capture rules, a file format, a rendered board. $29 once, no server to run.