MCP (Model Context Protocol)

MCP (Model Context Protocol) is an open source standard for AI-tool integrations. MCP servers give Claude Code access to external tools, databases, and APIs. This allows Claude to read and act on systems like issue trackers or monitoring dashboards directly, eliminating the need to copy data into chat.

To connect an MCP server, Claude Code uses specific commands. For remote HTTP servers, the command is claude mcp add --transport http <name> <url>. Similarly, claude mcp add --transport sse <name> <url> is used for Server-Sent Events (SSE). For a local stdio server, the command is claude mcp add [options] <name> -- <command> [args...], where the double dash (--) separates Claude's options like --transport, --env, or --scope from the actual server command. WebSocket is also supported as a transport, though it does not support OAuth or the --transport flag.

Configuration for MCP servers lives in .mcp.json, ~/.claude.json, or can be added via claude mcp add-json. Within JSON configuration, the type field accepts streamable-http as an alias for http. It is important to note that a JSON entry with a url but no type field is a configuration error, as Claude Code interprets such an entry as a stdio server. Beyond providing tool access, an MCP server can also act as a channel that pushes messages into a session.

MCP servers are fundamental for extending Claude Code's capabilities, much like how hooks or plugin might integrate functionality. For example, the MCP Builder skill helps in creating these servers, while Trace MCP routes code exploration to a code-graph MCP server. Understanding MCP is key to leveraging Claude Code's interaction with external systems, distinguishing it from a standalone agent-skill.