Claude Skills vs MCP: What's the Difference and When to Use Which

MCP connects Claude to external tools and data; a skill teaches Claude a procedure for using them. Skills do not replace MCP. Use a skill for know-how and an MCP server for live access. Claude skills and MCP servers are two ways to extend an agent. They differ in purpose, context cost, setup and safety. The table sets them side by side, and the sections below decide by goal.

Checked against Anthropic and MCP docs on 2026-10-20.

What is the difference between Claude skills and MCP?

The difference between Claude skills and MCP is purpose. MCP connects Claude to external services and data, such as GitHub or a database. A Claude skill gives Claude procedural knowledge, and MCP supplies the tools it uses.

Anthropic's Help Center puts the split in one line:

MCP connections give Claude access to tools, while skills teach Claude how to use those tools effectively.

The Model Context Protocol is an open standard for linking AI applications to outside systems. Its own docs call MCP "a USB-C port for AI applications." An MCP server exposes tools, resources and prompts that any MCP client can call. An Agent Skill is a folder with a SKILL.md file, published as an open standard at agentskills.io. Claude reads the skill's name and description first and loads the rest only when a task matches.

Skills are instructions by design. Their gain is progressive disclosure: a skill costs about 100 tokens until Claude needs it. Our guide to what Agent Skills are covers the format. The page on how Claude picks a skill covers selection and real examples.

Do Claude skills replace MCP?

No, Claude skills do not replace MCP; Anthropic describes the two as complementary. A CLI plus a skill covers some jobs an MCP server once did. Live, authenticated access to another system still needs MCP.

The "MCP is dead" claim runs through forum threads and Hacker News comments. Three positions frame the debate:

• Anthropic's engineering post says skills complement MCP servers by teaching agents complex workflows with external tools.
• Simon Willison, in an opinion review on his blog dated 2025-10-16, takes the blunter view.
• Anthropic's Help Center keeps MCP for access and skills for method, so each covers a gap in the other.

Willison wrote: "Almost everything I might achieve with an MCP can be handled by a CLI tool instead."

His point holds for agents with a shell and fails where none exists.

When does a CLI plus a skill beat an MCP server?

A CLI plus a skill beats an MCP server when the agent has a shell and a ready command. The skill documents the commands, and no server runs. Three goals decide the pick:

• Shell access with an existing CLI points to a CLI plus a skill.
• Several clients needing the same tool point to an MCP server.
• Chat in the Claude app, with no shell, points to an MCP connector.

How do skills and MCP servers use context and tokens?

Skills and MCP servers both defer most of their context. A skill keeps about 100 tokens of name and description loaded and reads its body only when triggered. Claude Code's tool search loads MCP tool names and server instructions first.

Skill files cost nothing until Claude reads them. Claude Code caps the whole skill listing at 1% of the context window. A large skill set overflows that budget. On the MCP side, Claude Code turns tool search on by default. Its docs state: "Only tool names and server instructions load at session start." Server instructions and upfront tool descriptions are cut at 4,096 characters. Deferred tool descriptions stop at 16,384. The MAX_MCP_OUTPUT_TOKENS variable changes the 25,000-token output limit.

Willison put the GitHub MCP server at "tens of thousands of tokens" in October 2025. That figure predates tool search, so it describes the older loading model. Neither side wins on context cost by default today.

When should you use a skill, and when an MCP server?

Use a skill when Claude already has the tools but needs your procedure, format or checklist. Use an MCP server when Claude needs live data or actions in another system. Use both when one workflow needs each.

A skill's script can call an API directly. MCP standardises the connection and the auth for every MCP client, so a team writes the integration once. Frontend Design is an instruction-only skill we reviewed. Its verdict is Recommended, 76, Moderate confidence, safety grade A. Our scan found two text files with no executable code, network calls or credentials.

The other half of the choice is the server itself. To compare servers before you connect one, read independent MCP server reviews on MCPVerdict, a separate site.

Can a Claude Code skill use MCP tools?

Yes. A Claude Code skill can use MCP tools by naming which to call and how to check results. Anthropic's partner skills for Notion, Figma and Atlassian work with their MCP connectors.

Anthropic's skills-and-MCP post gives a worked example. Notion's Meeting Intelligence skill defines which pages to search and how to structure the output. The Notion MCP connection does the searching, reading and page creation. Claude Code plugins bundle skills and MCP servers in one install. A vague description stops the skill from loading at all. Read why a skill does not fire before you blame the server. Install paths and folder rules sit in our Claude Code skill setup guide.

A skill that teaches an MCP server's tools

A skill that teaches an MCP server's tools fails whenever the connection fails. The Figma skill from OpenAI's skills repository shows the pattern. It runs a fetch-context, fetch-screenshot, adapt-to-project loop over the Figma MCP server. The skill has no scripts of its own. Its verdict, our Figma skill verdict, is Conditional, 58, Moderate confidence, safety grade C. Our evidence verifies it only in Codex with the Figma MCP connected. Its negative evidence centres on the MCP link: OAuth invalid_redirect_uri errors, tools not injected, and 401s.

Pre-approving MCP tools in a Claude Code skill

Pre-approving MCP tools in a Claude Code skill uses the allowed-tools field. Claude Code names every MCP tool mcp__<server>__<tool>. The grant covers the turn that invokes the skill, and other tools keep your normal permission settings. This pattern suits custom skills built on internal MCP services:

---
name: weekly-crm-report
description: Builds the weekly sales report from the internal CRM. Use when asked for the weekly report.
allowed-tools: mcp__crm__search_deals mcp__crm__get_account
---
Search open deals first, then fetch each account and fill the report template.

Where skill and MCP instructions overlap

Skill and MCP instructions overlap at the edges, and Anthropic draws the line in three places:

• MCP server instructions cover how to use the server and its tools correctly.
• Skill instructions cover how to use those tools for one process or across several servers.
• Duplicate rules in both places waste context and drift apart over time.

Are skills or MCP servers safer?

Neither skills nor MCP servers are safe by default. Skills can carry malicious instructions or scripts, and MCP servers can relay prompt injection from fetched content. Anthropic's skill scanning covers uploaded skills and plugins, but not MCP servers.

OWASP's Agentic Skills Top 10 separates the layers. Its line reads: "MCP = how the model talks to tools; AST10 = what those tools actually do." Two outside studies sized skill risk. An arXiv study (2601.10338) found a vulnerability in 26.1% of 31,132 skills. Snyk found critical issues in 13.4% of 3,984 skills, or 534.

Our Figma scan shows the MCP risk in practice. The skill gives no guidance on treating text inside Figma designs as untrusted input.

Every verdict on this site carries an A to F safety grade from reading the skill's files. Start with how to vet a skill before you install it. Our methodology page explains how the safety grade is set.

What do our verdicts show about skills that need an MCP server?

The skills below each list an MCP server they need or recommend. The two Figma skills require one and score Conditional. Using Neon treats MCP as optional and scores Recommended.

Using Neon is the one row that works without MCP. Its scan found five Markdown files and no executable code. The workflows still assume a Neon account, a DATABASE_URL and CLI or MCP tooling that runs SQL.

All three skills in the table carry grade C, which marks skills that need credentials or broad permissions. That holds for three skills and sets no rule.

For the server half of these pairs, see MCPVerdict's server verdicts. For skills without that dependency, see Claude Code skills ranked by verdict.

The choice between skills and MCP ends there. The sections below cover the wider vocabulary, the Claude app's wording and other agents.

How do skills, MCP, tools and agents fit together?

Skills, MCP, tools and agents fit together as layers. An agent is the model running a loop, and tools are functions it calls. MCP servers supply tools, and skills tell the agent when and how to use them.

• An agent is the model plus the loop that plans, calls tools and reads results.
• A tool is a function with a schema that the agent calls.
• An MCP server is a process that supplies tools, resources and prompts over the protocol.
• A skill is a SKILL.md folder that Claude loads when its description matches the task.

To find a skill, browse every skill verdict.

Are Claude connectors the same as MCP or skills?

Claude connectors are MCP servers as people see them in the Claude app. A connector gives Claude access to a service. Skills are separate from connectors, and a skill teaches Claude a workflow that uses one.

Anthropic's connector docs say it directly:

An MCP server gives Claude access to your product or data, and people who use Claude see it as a connector.

Partner skills for Notion, Figma and Atlassian are built to pair with those connectors. The app setup steps show how to turn on skills in the Claude app.

Do skills and MCP work the same way in Codex and Cursor?

Yes. Skills and MCP work the same way in Codex and Cursor: both read SKILL.md and connect MCP servers separately. Of the skills we have reviewed, 7 are verified in Codex.

A Codex skill can declare its MCP dependencies in agents/openai.yaml, per OpenAI's skills docs. Cursor's docs say it reads SKILL.md folders from its skills directories. Cursor support is thinner in our reviews: 1 skill is verified there. Verified means a dated evidence item shows the skill working in that agent. For Claude itself, see the Claude skills worth installing.

Common questions

Can an MCP server expose skills?

A proposal says yes. SEP-2640, posted on 2026-06-18, proposes serving skills as MCP resources under skill:// addresses.

Can you add skills to Claude through MCP?

Yes, in Claude organisations. Those skills follow the MCP trust model and skip Anthropic's skill scanning.

Is a skill an MCP server?

No. A skill is a folder of files with a SKILL.md, and no server process runs.

Can one skill use several MCP servers?

Yes. Name each tool in the skill and pre-approve it with allowed-tools.

Should you build a skill or an MCP server for an internal API?

Build an MCP server for live calls from many clients. Write a skill when Claude already reaches the API and needs your procedure.

Are Claude skills any good?

Many are. Of the skills we have reviewed, 10 are Recommended and 10 are Recommended with caveats; see every skill we have reviewed.