MCP in Claude Code

What is MCP?

• Open standard that connects cloud code to external tools and data sources
• Automatically determines when to use tools to better answer queries
• Brings context from databases, productivity apps, and repositories into cloud code

Tools in Agentic AI

• Tools enable agents (like cloud code) to perform actions, not just return text
• Different from traditional AI that only provides text output
• Examples: Linear MCP server for project management, Context 7 for documentation

Types of MCP Servers

• HTTP servers: Remote services hosted by providers, connect over network
• STDIO servers: Local processes running on your machine

Adding and Managing Servers

• Add servers with: cloud MCP add command
• Manage with /MCP command inside cloud code session to view status and disable servers
• Browse hundreds of connectors at cloud. com/connectors

Scoping MCP Servers

• Local: Available only in current project for you
• User: Available across all your projects
• Project: Uses . mcp. json file checked into version control—team gets same servers automatically

Context Window Efficiency

• MCP servers add tool definitions to context even when unused
• Disable unused servers to preserve context space
• CLI tools (e. g. , gh, aws) are more context-efficient than MCP equivalents
• Consider using skills instead—they load into context only when needed
• If MCP tools exceed 10% of context, Claude automatically switches to tool search mode

Takeaways

• MCP is an open standard that connects AI agents to external tools and data sources, automatically determining when to use them to enhance responses beyond text-only output.

• MCP servers come in two types: HTTP servers (remote, network-based) and STDIO servers (local processes)—manage them with cloud MCP add and /MCP commands.

• Context window efficiency matters: MCP tool definitions consume context space even when unused; disable unnecessary servers, prefer CLI tools, or use skills (which load only when needed) to preserve context.

• Scope servers appropriately: Local (current project only), User (all your projects), or Project (shared via . mcp. json in version control for team consistency).