Claude Code connects to MCP servers for external tools and data. Use it when the agent needs authoritative context or must interact with another system through a structured interface.

Add and Inspect Servers

Use the command line interface:

claude mcp add
claude mcp list
claude mcp get
claude mcp remove

For OAuth-backed servers, use:

claude mcp login
claude mcp logout

Project-scoped servers can live in .mcp.json. Use project scope when the server is part of the repository workflow. Use user scope for personal tools.

Prefer HTTP for New Remote Servers

Claude Code supports stdio, HTTP, and SSE transports. For new remote servers, prefer HTTP where possible. SSE remains supported for compatibility, but the current documentation marks it as deprecated.

Permissions

MCP tools run through Claude Code’s permission model. Keep write tools behind ask rules unless the environment is isolated and the action is low risk.

Prompt important tool use explicitly:

Use the GitHub MCP server only to read the pull request checks and comments.
Do not post comments or update labels.

Security Review

Before adding a server, answer:

• Which external systems can it access?
• Which credentials does it receive?
• Can it write to shared systems?
• Does it expose logs, prompts, or tool arguments?
• Can project configuration add or replace the server?

MCP is a capability boundary. Review it like one.