Claude Skills vs MCP
Updated September 17, 2026
Claude Skills and MCP get compared because both extend what Claude can do, but they extend different things. A skill changes what Claude knows: it is a folder of instructions loaded when a task matches. MCP, the Model Context Protocol, changes what Claude can reach: it is an open standard for connecting to outside tools and data through servers. Most real setups use both, and this guide explains where the line falls.
The short answer
Use a skill to capture a procedure, a format, or domain knowledge that Claude should apply consistently. Use MCP when Claude needs to talk to a system it cannot otherwise see: a database, an issue tracker, a SaaS API, a browser. When a procedure involves such a system, write a skill that tells Claude how to use the MCP tools well. They are layers, not rivals.
What a skill is
A skill is a directory containing a SKILL.md file: YAML frontmatter with a name and a description, then Markdown instructions, optionally with scripts and reference files alongside. Claude loads every skill's name and description at startup and reads the full file only when a request matches, so a hundred installed skills cost about a hundred tokens each until one is used. Skills need no server and no network. A bundled script can run locally on data Claude already has, but a skill cannot by itself reach a remote service. How to create a Claude skill walks through building one.
What MCP is
MCP is an open standard for AI-tool integrations. An MCP server is a running program, local or remote, that exposes three kinds of capability: tools Claude can call, resources it can read, and prompts that appear as slash commands. Claude Code adds one with a command such as claude mcp add --transport http notion https://mcp.notion.com/mcp, at local, project, or user scope; project-scope servers are stored in .mcp.json and shared through version control. Servers that need OAuth are authorised through /mcp. Once connected, the server's tool definitions are available to Claude for the session.
Side by side
| Skill | MCP server | |
|---|---|---|
| What it changes | What Claude knows: procedures, rules, formats | What Claude can reach: external tools, data, services |
| Form | A folder with SKILL.md, optional scripts and references | A running process speaking the MCP protocol, local (stdio) or remote (HTTP) |
| When it loads | Name and description at startup; full file when triggered or invoked | Tool definitions when the server connects |
| Needs a network? | No; scripts run locally | Usually; remote servers by definition |
| Setup | Create a file; commit or zip it | Add the server, authenticate, approve it per project |
| Sharing | Git, plugins, zip upload to claude.ai, Skills API | .mcp.json in the repo, plugins, or each user adds it |
| Typical use | Release notes, review checklists, report templates, migrations | Reading Jira, querying Postgres, posting to Slack, browsing |
| Main risk | A malicious skill instructs Claude to misuse tools it already has | A server that fetches external content can carry prompt injection |
When you need both
The common case is a procedure that touches a system. “Triage new bug reports” needs MCP to read the tracker and a skill to say how triage is done here: which labels, which severity rules, what a good summary looks like. Without the skill Claude has access and no method; without MCP it has method and no access.
When a skill refers to MCP tools, Anthropic's guide says to use the fully qualified name, ServerName:tool_name, such as GitHub:create_issue, because with several servers connected an unqualified name may not be found. In Claude Code, a skill's allowed-tools list can pre-approve specific MCP tools using their full names, so a well-scoped skill runs without a prompt for each call.
Security: treat both like software
Anthropic's guidance for skills is to use only ones you wrote or obtained from Anthropic, and to audit anything else: read SKILL.md and every bundled script, and look for network calls or file access that do not match the stated purpose. A skill can direct Claude to invoke tools in ways you did not intend. For MCP, the Claude Code documentation is equally direct: verify you trust each server before connecting it, since servers that fetch external content can expose you to prompt injection. Project-scoped servers from a repository need your approval before they are used.
Choosing, in one list
- Repeating the same instructions across chats: a skill.
- Claude cannot see the data it needs: an MCP server.
- Claude has the tool but uses it badly: a skill that explains the tool.
- A deterministic transformation on local files: a skill with a script.
- Live, changing information from another system: MCP.
- Both of the above in one workflow: a skill that names the MCP tools it should use.
Common questions
- Is a skill a replacement for MCP?
- No. A skill cannot reach an external system at runtime; MCP exists for that. A skill can tell Claude how and when to use the tools an MCP server provides, which is why the two are usually combined.
- Which costs more context?
- A skill costs roughly a hundred tokens for its name and description until it is triggered, then its body. MCP tool definitions are present once the server is connected. In both cases, keep what you install to what you use.
- Can a skill call an MCP tool?
- A skill's instructions can tell Claude to use one, referenced as ServerName:tool_name. The server must be connected in the session; the skill does not connect it.
- Do skills work without Claude Code?
- Yes. Skills follow the open Agent Skills specification and work in claude.ai and through the Claude API, and other agents have adopted the format. MCP is likewise an open standard used by many clients beyond Claude.
- Which should I set up first?
- The one that removes your bigger friction. If you keep re-explaining a process, write a skill in ten minutes. If Claude keeps asking you to paste data it should be able to read, connect an MCP server for that source.
Try it yourself
These live under Developer tools, alongside the rest of the claude skills & skill.md guides on the guides page.