← All guides

Locus product guide · Updated

Claude Skills vs MCP: what is the difference?

MCP gives an agent new tools and data: a server exposes functions the agent can call, often in another system. A skill teaches the agent how to do a task: a SKILL.md file of instructions, plus optional scripts, loaded only when relevant. An unused skill costs about 100 tokens of context. Anthropic’s help center says the two work together: MCP for access, skills for how to use that access well.

How do Claude Skills and MCP compare?

They solve different problems, so the comparison is about fit. Facts below come from the Agent Skills spec, Anthropic’s docs and the MCP 2026-07-28 spec.

How do Claude Skills and MCP compare?
QuestionAgent SkillsMCP
What it addsKnow-how: steps, rules, examples, scriptsAccess: tools, resources and prompts from a server
FormatA folder with SKILL.md and optional filesJSON-RPC 2.0 messages between a client and a server
Where it runsInside the agent’s own environment, using its file and code toolsIn a separate process, on your machine or on the internet
Context costAbout 100 tokens per skill until used, under 5,000 tokens when loadedTool names, descriptions and schemas, unless the client defers them with tool search
CredentialsNone built in. Scripts use whatever the environment has.OAuth 2.1 for HTTP servers, environment variables for stdio
Live data and actionsOnly through tools the agent already hasYes, that is its main job
Who keeps it currentThe skill author, by editing filesThe server operator, by updating the server
Open standardagentskills.iomodelcontextprotocol.io, current version 2026-07-28

When is a skill the better choice?

Pick a skill when the agent already has the access it needs but does the task badly or inconsistently.

  • Team conventions: commit messages, report formats, brand rules.
  • Multi-step procedures the agent should follow the same way every time.
  • Reference material the agent should read only when relevant, such as a database schema.
  • Deterministic steps you can script, such as validating a file.
  • No new credentials. A skill should not be the place you store an API key.

When do I need an MCP server instead?

Pick MCP when the agent needs something it cannot reach from its own environment. On the Claude API, skills run with no network access at all, so any outside call needs a tool.

  • Live data from another system, such as search results or a CRM record.
  • Actions in another system, such as creating a ticket or sending a message.
  • Credentials that should stay out of the agent’s environment and be scoped per user with OAuth.
  • One integration shared by many clients: Claude, ChatGPT, VS Code and Cursor all speak MCP.
  • Central rate limits, logs and revocation on the server side.

How do skills and MCP work together?

The Claude help center puts it simply: MCP connections give Claude access to tools, and skills teach Claude how to use those tools. Partner skills from Notion, Figma and Atlassian are built to work with their MCP connectors.

The two are also converging. The MCP Skills extension, io.modelcontextprotocol/skills, lets a server publish skills that a client lists and reads as MCP resources, with a digest check on every file. Its proposal, SEP-2640, is final, but host and SDK support is still rolling out.

Locus Pro ships both. The Claude Code plugin installs the hosted Locus MCP server and three skills. The locus skill tells Claude when to use Locus, for example before it scrapes a site with curl, and how to search, describe and execute a tool. The MCP server does the work and bills each call.

What does each approach cost?

A skill has no fee. It costs context: about 100 tokens while idle and up to a few thousand when loaded. An MCP server costs context for its tool list, and its tools may charge per call.

Big tool lists add up. The Locus Pro compact surface lists four tools, search, describe, execute and get_result, instead of one tool per provider. Search and describe are free. Paid calls cost the catalog price: Firecrawl search is 1.7 credits and Brave web search is 30.3 credits, where 1,000 credits equal $1. Prices checked November 10, 2026.

Should I build a skill, an MCP server or both?

Answer two questions: does the agent lack access, and does it lack know-how?

  • Lacks access only: add an MCP server.
  • Lacks know-how only: write a skill.
  • Lacks both: add the server and a skill that explains when and how to use it.
  • Your own code calls a known API: neither. Call the API directly.

Frequently asked questions

Do skills replace MCP?

No. A skill cannot reach a system the agent has no access to. It can only tell the agent how to use the tools it has, which often include MCP tools.

Can a skill tell the agent to use an MCP tool?

Yes. A skill’s instructions can name the tools to call and in what order. The allowed-tools field can also pre-approve tools for the skill, though support for it varies by agent.

Can an MCP server ship skills?

Yes, through the MCP Skills extension. The server lists skills and serves their files as resources, and the host checks each file’s SHA-256 digest before loading it. Host support is still being built, so check your client.

Which uses less context, a skill or an MCP server?

An idle skill uses about 100 tokens. An MCP server adds every tool definition it lists, unless the client loads tools on demand. Servers with a small tool surface keep that cost down.

Implementation references

Use these first-party references for current request contracts and account requirements. Tool availability, prices, and negotiated terms can change.

Sources

Third-party details were checked on November 10, 2026. Vendors change pricing and features often, so confirm on their site before you decide.

Related guides