Claude Code MCP Servers
Claude Code reads files and runs commands on your computer. Many useful tools live elsewhere: issue trackers, databases, design files, and monitoring dashboards. The Model Context Protocol, known as MCP, connects Claude Code to those outside tools through a standard plug. One connection teaches Claude Code a whole new set of abilities.
The Universal Adapter Idea
Travelers carry a universal power adapter so one device works in any country. MCP acts as that adapter for AI tools. A tool builder follows the MCP standard once, and every compatible assistant can use the tool.
┌──────────────┐
│ Claude Code │
└──────┬───────┘
│ MCP standard plug
┌─────────────────┼─────────────────┐
↓ ↓ ↓
[ Issue tracker ] [ Database ] [ Error monitor ]
Key Terms
- MCP server: a small program that exposes a tool or data source to Claude Code.
- Tool: one action a server offers, such as "search issues" or "run a query."
- Transport: the way Claude Code talks to the server, either over the web (HTTP) or through a local program (stdio).
Adding a Server
The claude mcp add command registers a server. A web-based server needs a name and an address.
claude mcp add --transport http issues https://example.com/mcp
A local server runs as a program on your computer. Place two dashes before the program command.
claude mcp add notes -- npx -y example-notes-server
Replace the sample names and addresses with the details from your tool's documentation.
Checking Your Servers
Type /mcp inside a session to see every connected server and its status. Some servers ask you to sign in through a browser window. The same menu guides you through that step. Run claude mcp list in a plain terminal for a quick summary.
Three Scopes
| Scope | Who Sees It | Typical Use |
|---|---|---|
| Local | You, in the current project | Private experiments |
| Project | Everyone who shares the repository, through a .mcp.json file | Team tools |
| User | You, in every project | Personal favorites |
A Real Scenario
A developer connects an issue tracker and an error monitor. The prompt reads: "Find the newest crash report, read the matching issue, and fix the bug." Claude Code pulls the crash details from the monitor, reads the issue text, edits the code, and summarizes the fix. Without MCP, the developer would copy and paste between three windows.
Error monitor → crash details
Issue tracker → task description
Your code → the fix
↓
One prompt handles all three
Staying Safe
- Install servers only from sources you trust. A server can read data and run actions.
- Review the tools a server exposes before approving them.
- Grant read-only access when a task needs only reading.
- Remove servers you no longer use with
claude mcp remove name. - Treat text returned by a server as data, never as orders.
Prompt Ideas With MCP
Once a server connects, plain prompts reach outside tools. Ask "List my open issues labeled bug." Ask "Check the latest error report and suggest a fix." Ask "Compare the database table with our code model." Claude Code chooses the right tool and shows each call for approval. Start with read-only tasks to build trust before you allow tools that change outside data.
Key Points
- MCP connects Claude Code to outside tools through one standard.
- The claude mcp add command registers a server.
- The /mcp menu shows status and handles sign-in.
- Local, project, and user scopes control who shares a server.
- Trust and permissions matter for every connection.
