The one sentence difference
MCP is capability. A skill is knowledge.
The Model Context Protocol connects an assistant to a running server that exposes tools: functions the model can call with arguments and get results back. The server can reach databases, APIs, and user accounts, so the assistant can answer questions about live state and perform real actions.
A Claude Skill is a folder with a SKILL.md: a description that tells the assistant when it applies, and a body of instructions it loads at that moment. No process, no network, no auth. It cannot fetch your user’s data. What it can do is make the assistant behave like someone who has actually read your documentation, because in effect it just did.
What each one is good at
MCP server strengths
- Live, account specific answers: “which of my users cost me money this month?”
- Actions with side effects: create a project, rotate a key, set a spending cap.
- Data that changes faster than any document: today’s spend, current margins, active alerts.
Claude Skill strengths
- Correctness before the account exists: the assistant writes your real SDK syntax instead of a plausible invention, which matters because assistants now write most integration code.
- Method, not just facts: how to run a cost estimate, what order to audit plans in, what not to do.
- Zero infrastructure: markdown in a public repo, installable with one command, nothing to keep up.
The decision in practice
Ask one question: where does the assistant fail without you?
- It answers wrongly about your product. Invented API calls, stale prices, wrong mental model. Ship a skill. This is a knowledge gap and a running server does nothing for it.
- It cannot answer at all without live data.Anything with the user’s own numbers in it. Ship MCP. No amount of markdown contains this month’s spend.
- Both failures exist. Which is most developer tools: users need correct integration first and live answers later. Ship both, and make the skill mention the MCP server so one install leads to the other.
Concretely for Weckr: the Weckr Claude Skills teach the assistant the two line integration, current model prices from a live LLM pricing feed, cost estimation, and plan margin audits. The Weckr MCP server then answers questions against your real dashboard data. Before and after, knowledge and capability.
Cost of ownership, honestly
The skill took us about a day to write well, and its only ongoing cost is keeping facts current, which we automated: the pricing tables defer to a live feed and a weekly watcher opens a pull request when provider rates move. The MCP server is real software: it holds API keys, talks to production, and has to fail safely. Neither is hard, but they are different classes of commitment, and the skill is the cheaper first step by a wide margin. How we built ours is in why your SDK should ship a Claude Skill.
FAQ
What is the difference between an MCP server and a Claude Skill?
An MCP server gives the assistant live access to data and actions: it is a running process that exposes tools the model can call, like query my dashboard or create a ticket. A Claude Skill gives the assistant know-how: it is a folder of markdown instructions the model loads when a task matches, like the correct syntax for an SDK or the right method for a cost audit. MCP is capability, a skill is knowledge.
When should a product ship an MCP server?
When the value is in the user reaching their own live data or performing actions through the assistant. If the useful question is which of my users cost me money this month, only a live connection can answer it, so that is MCP territory. MCP also fits write actions: creating resources, changing settings, triggering jobs.
When should a product ship a Claude Skill instead?
When the failure mode is the assistant getting your product wrong: inventing SDK syntax, using stale prices, or applying the wrong method. A skill fixes that with zero infrastructure, since it is just markdown in a repo. It works before the user has an account, which makes it the better top of funnel artifact: the assistant integrates your product correctly for someone who has never opened your docs.
Can you ship both an MCP server and a skill?
Yes, and for developer tools it is often the right answer because they cover different moments. The skill handles before and during integration: correct code, correct concepts, correct math. The MCP server handles after: querying the live data the integration produces. Weckr ships both, and they reference each other rather than compete.
What does each cost to build and run?
A skill is close to free: markdown files, a small plugin manifest, no servers, no auth, no uptime. The main cost is keeping its facts current. An MCP server is software you run and maintain: it needs auth to the user account, error handling, and versioning. Rule of thumb: a skill is a docs problem, an MCP server is a product surface.
Keep reading
See both halves working
Install the Weckr Claude Skills to watch the knowledge half write correct integration code, then wire the MCP server to query your own cost data. The product both sides point at is on the live demo, and the wider playbook is the AI cost and margin guide.