For the last four levels, everything Claude knew lived in the repo: code, CLAUDE.md, skills, subagents. But real work lives elsewhere too: issues on GitHub, errors in Sentry, data in a database, designs in Figma. If you're copying by hand from those tabs into the chat, this level is for you. The first lesson explains what MCP is, how it differs from a skill, and connects a server that needs no account.
What MCP is
MCP (Model Context Protocol) is an open-source standard for connecting AI to outside tools. An MCP server sits in the middle: on one side it talks to a service (GitHub, a database, a browser…), on the other it hands Claude Code tools for using that service.
The sign that you need MCP is clear: you keep copying data from another tool into the chat. Open an issue, copy it, paste it to Claude. Open a dashboard, copy the numbers, paste again. Once a server is connected, Claude reads and acts on that system directly instead of working from what you paste.
Some examples from the docs:
- "Add the feature described in issue ENG-4521 and create a PR on GitHub."
- "Check Sentry to see how that feature is failing."
- "Find 10 users who used this feature in our PostgreSQL database."
A server can run on your machine or be a hosted service on the internet. Because it is an open standard, a server built on MCP works with other AI apps that support MCP, not only Claude Code.
What a server provides
A server can give Claude Code three things:
- Tools: actions Claude can call, such as searching issues, running a query or opening a web page. Tool names look like
mcp__<server>__<tool>, for examplemcp__github__list_issues. You'll meet this form when you set permissions. - Resources: data you pull into a prompt with
@, like mentioning a file:@github:issue://123. Type@to see them in the suggestions. - Prompts: ready-made prompts from the server, which show up as slash commands such as
/mcp__github__pr_review 456.
Every server has tools; resources and prompts depend on the server.
There are two main ways to connect:
- HTTP: the server lives online and you give it a URL. The docs recommend this for remote servers. The older SSE style is being phased out.
- stdio: the server is a program running on your machine, such as
npx some-mcp-server. A good fit for tools that need direct access to your machine.
How MCP differs from a skill
Both "extend Claude", so they're easy to mix up. The docs draw a clean line:
| MCP | Skill | |
|---|---|---|
| What it is | A protocol for connecting to outside services | Knowledge, workflows, reference material |
| Gives Claude | Tools and data access | A way of working |
| Examples | Slack integration, database queries, browser control | Review checklist, deploy workflow, API conventions |
In short: MCP gives Claude hands that can reach other systems. A skill gives Claude the know-how to use those hands well.
They don't replace each other; they work together. Take a database: an MCP server handles the connection and sign-in, while a skill records your team's schema and the queries you run most. Without MCP, Claude can't reach the database. Without the skill, it can reach it but has to guess which table holds what.
As for the order to add things to a project, the docs suggest following the signs: Claude gets a convention wrong twice, fix CLAUDE.md; you paste the same procedure for the third time, write a skill; you keep copying data from a tab Claude can't see, connect MCP.
Your first server: the Claude Code docs
The docs offer a hosted server that needs no sign-in and searches the Claude Code documentation. Use it to learn the whole flow. Run this in your terminal, not inside a claude session:
claude mcp add --transport http claude-code-docs https://code.claude.com/docs/mcpclaude-code-docs is a name you choose. Claude Code uses it to label the server's tools and in management commands. The command prints "Added HTTP MCP server … to local config" and the file it changed.
Check the connection:
claude mcp listThe server's line should show ✔ Connected. When we tried it, the list also showed our claude.ai account's connectors (Google Drive, Gmail…), because Claude Code loads those too. Other statuses: ! Needs authentication (needs a sign-in, lesson 5.2) and ✘ Failed to connect (the server didn't answer).
Try it: start claude and ask:
Use the claude-code-docs server to look up what MCP_TIMEOUT doesNaming the server makes sure the answer goes through it; normally Claude picks the right tools on its own. You may be asked for permission the first time; allow it. In the output, the tool call is labelled with the server's name: that is your proof the answer came from the server and not from Claude's memory. In our run, Claude answered correctly: the time to wait for a server to start, in milliseconds, 30 seconds by default.
Remove it when done: claude mcp remove claude-code-docs. Inside a session, type /mcp to see the status of your servers.
By default claude mcp add saves the server at local scope: just you, just this project, in ~/.claude.json. Sharing with your team through .mcp.json is the topic of lesson 5.2.
The context cost
Every connected server takes up a little context. Claude Code keeps that small with tool search, which is on by default: at the start of a session it loads only the tool names and the server's instructions, and loads a tool's full description only when Claude needs it. We saw exactly that in our test: Claude took a ToolSearch step to load the docs server's two tools before using them.
Results have limits too. Output over 25,000 tokens (the default, set with MAX_MCP_OUTPUT_TOKENS) or over 50,000 characters is saved to a file, and Claude reads that file when it needs to. One of our searches returned 80 KB and was saved to a file rather than flooding the conversation.
Even so, a server you don't use still costs space in every session. Remove servers you no longer use.
Trust a server before you connect it
Connecting a server lets Claude use someone else's tools on real data. The docs are explicit: check that you trust a server before connecting it. A server that fetches outside content (web pages, issues written by strangers) can carry prompt injection: text written on purpose to trick Claude into doing something you didn't want.
Permissions for MCP tools work like any other tool, by name:
mcp__claude-code-docsmatches every tool from that server;mcp__github__list_issuesmatches one tool.
A safe habit: pre-approve the read-only tools and let Claude Code ask every time for tools that write or delete. Tokens, permissions and safety are the topic of lesson 5.3. Next, lesson 5.2 connects GitHub and shares server configuration with the whole team through .mcp.json.