Skills vs MCP: when to use each
Agent skills and MCP servers get lumped together, but they solve different problems — and the best setups use both. Skills are procedural know-how that loads on demand; MCP servers are live connections to tools and data. Here is the practical breakdown.
Which do you need? Answer 2 quick questions
1. Does the task need live data or actions in an external system (database, API, SaaS)?
2. Is it a repeatable how-to workflow the agent should follow?
3. (Optional) Should it work across multiple agents — Cursor, Codex, Gemini CLI?
Answer the first two questions to see a recommendation.
Side by side
| Aspect | Skills | MCP servers |
|---|---|---|
| What it is | A SKILL.md folder of instructions (+ optional scripts) | A protocol server exposing tools, resources and prompts |
| Best for | Procedural know-how — how to do a task well | Connections — reaching external data and systems |
| Activation | Loaded on demand when relevant | Tools called explicitly during a task |
| Token cost | A few dozen tokens until loaded | Tool schemas can cost thousands of tokens upfront |
| Setup | Drop a folder into ~/.claude/skills/ | Install + configure a server, manage credentials |
| Portability | Same files across Claude, Cursor, Codex, Gemini CLI | Works with any MCP client; per-client config paths |
Use a skill when…
- You want the agent to follow a repeatable workflow (TDD, debugging, code review).
- The task is about how to do something, not about fetching live data.
- You care about token budget and want capabilities to load only when needed.
- You want the same behaviour across Claude Code, Cursor, Codex and Gemini CLI.
Use an MCP server when…
- The agent needs live data from a database, API or SaaS (Postgres, GitHub, Slack).
- You need to perform actions in an external system, not just reason about a task.
- Access must be authenticated and scoped with real credentials.
- You are integrating a tool many agents will call repeatedly.
Best of both: combine them
The strongest agents pair a skill with the MCP server it needs. The webapp-testing skill plus a Playwright MCP server. The requesting-code-review skill plus the GitHub MCP server. Assemble a combined stack in the Builder.
Context budget: skills vs MCP
MCP servers advertise their tool schemas upfront; skills cost only a few dozen tokens until loaded. Estimate the difference for your setup.
Same capabilities as skills would free roughly 31,760 tokens of context before any work starts — about 16% of your window.
Rough estimate. Skills cost ~40 tokens of metadata each until loaded on demand; MCP servers advertise their tool schemas upfront — GitHub’s official server alone can run into tens of thousands of tokens.
FAQ — Skills vs MCP
Are skills replacing MCP?
No. They solve different problems and work best together. Skills give an agent procedural know-how that loads on demand; MCP connects the agent to external tools and data. A good setup uses skills for the workflow and MCP for the connections.
Why are skills more token-efficient than MCP?
A skill only exposes a short description (a few dozen tokens) until the agent decides it is relevant, then loads the full instructions. MCP servers advertise their tool schemas upfront, which can consume thousands of tokens of context before any work happens.
Can I use skills and MCP at the same time?
Yes — that is the recommended setup. For example, pair the webapp-testing skill with a Playwright MCP server, or the requesting-code-review skill with the GitHub MCP server. The skill guides the workflow; the server provides the live access.
Do skills work outside Claude?
Yes. Anthropic opened the SKILL.md format as a standard, and it is read by Cursor, Codex CLI, Gemini CLI and many other agents — the same folder works across them.