What you’ll know by the end of this check
- What MCP actually is (not marketing — the wire protocol)
- How to add, scope, and manage MCP servers
- The context-cost rule: when to reach for MCP, when to use a CLI instead, when to prefer a Skill
The shortest possible answer
Model Context Protocol (MCP) is an open standard for connecting Claude Code to external tools and data sources — Linear, Jira, Snowflake, your internal APIs, docs servers like Context7. Claude sees them as tools and decides when to call them.
A lot of your actual work context lives outside your codebase. MCP is the bridge.
Two shapes of MCP server
| Type | Runs where | Example |
|---|---|---|
| HTTP | Remote, hosted by the vendor | Linear MCP, Atlassian MCP |
| Stdio | Local process on your machine | A Python script, a custom internal server |
Add either with:
claude mcp add <name>Manage with /mcp inside a Claude Code session. You’ll see which servers are connected, their status, and the tools they expose. You can disable any server you’re not using.
Scoping: who gets what
| Scope | Where it lives | Who sees it |
|---|---|---|
| Local | This project only, your machine | You |
| User | All your projects | You |
| Project | .mcp.json checked into the repo | Everyone on the team |
Project scope is the superpower. Check a .mcp.json into the repo and every teammate gets the same MCP servers automatically. No “wait, did you add Linear?” drift.
The context-cost rule
MCP servers load tool definitions into your context window the moment they connect — even if you never use them. If you have five servers connected but only use one, the other four are taxing every prompt.
Three rules:
- If a tool has a CLI equivalent, use the CLI.
ghfor GitHub,wranglerfor Cloudflare,awsfor AWS. CLI calls don’t add persistent tool definitions. MCP does. - If you want a capability but not a full server, use a Skill. Skills load a name + description into context and pull the full body only when needed. MCP is always-on; Skills are on-demand.
- If your MCP tools exceed 10% of your context window, Claude Code auto-switches to tool-search mode. Discovery becomes less reliable. Lean defaults beat fat ones.
When MCP is the right answer
- The tool has no CLI equivalent (most internal APIs).
- You need structured data Claude will reason about across multiple steps.
- The data source is genuinely external and stateful (Linear issues, Jira tickets, a Postgres schema).
When it’s the wrong answer: “let’s add an MCP for GitHub” when gh already does the job.
Things to try right now (15 minutes)
- Run
/mcpin an active Claude Code session. List every server that’s currently connected. - Disable any server you haven’t used in this session.
- Add one MCP server you’ll actually use — Context7 for up-to-date library docs is a safe first choice:
claude mcp add context7. - Ask Claude a question that requires that server. Confirm it fires.
The canonical version
Full official lesson is at anthropic.skilljar.com/claude-code-101/469797.
Ready to verify this check?
You’ve added at least one MCP server and used it. You can name the three scopes. You know the CLI > MCP > Skill tradeoff. Mark it cleared.