CHAPTERS
What MCP is and why it matters for Claude Code context
Defines Model Context Protocol (MCP) as an open standard that connects Claude Code to external tools and data sources. Frames the core value: pulling in relevant context that lives outside the code editor (databases, apps, public repos) so Claude can answer and act more accurately.
- •MCP is an open standard for connecting Claude Code to tools and data
- •Claude can decide when to use tools automatically based on your request
- •Most useful context often lives outside the current workspace (DBs, apps, repos)
- •MCP is positioned as the bridge that brings that external context in
Tools in agentic AI: actions vs. text-only outputs
Explains the foundational concept of “tools” for agentic AI systems like Claude Code. Emphasizes that tools enable actions and retrieval, not just generating text responses.
- •Tools let Claude Code perform actions to complete tasks
- •Agentic workflows differ from text-only AI output
- •Tool usage improves task completion by fetching/acting on real systems
Real examples of MCP servers (Linear + dependency docs)
Gives concrete examples of how MCP servers expand Claude Code’s understanding of your environment. Highlights using Linear for issue context and a documentation connector for up-to-date dependency information.
- •Linear MCP server can surface team-specific issue details
- •Context7 MCP server can provide up-to-date dependency documentation
- •MCP servers tailor Claude Code to your stack and workflows
Finding connectors and adding MCP servers
Points to the broader ecosystem of available connectors and how to install them. Introduces the command used to add MCP servers to Claude Code.
- •Hundreds of connectors available at claude.com/connectors
- •Add MCP servers via the `claude mcp add` command
- •MCP servers are modular connectors to tools/data sources
Server types: HTTP (remote) vs STDIO (local)
Breaks down the two major MCP server deployment models. Clarifies when you’d use remote network-hosted servers versus local processes running on your machine.
- •HTTP servers connect to remote services hosted by providers
- •STDIO servers run locally as processes on your machine
- •Choosing type depends on whether the service is remote or local
Managing MCP servers during a session with /MCP
Shows how to inspect and control MCP server connections within Claude Code. Focuses on checking status and disabling servers you don’t want active.
- •Use `/mcp` in-session to see connected servers and status
- •Disable servers you don’t want Claude to use
- •Server management is part of keeping setups clean and intentional
Scoping servers: local, user-wide, and project (.mcp.json)
Explains how MCP configurations can be applied at different scopes. Highlights project scoping via a checked-in config file so teammates get the same setup automatically.
- •Local scope: only available in the current project for you
- •User scope: available across all projects
- •Project scope: committed `.mcp.json` shares servers across the team
- •Version-controlled MCP config standardizes tooling across a codebase
Context-window tradeoffs: MCP tool definitions consume space
Warns that MCP servers add persistent tool definitions to the context window even when unused. Advises auditing and disabling inactive servers to preserve context for actual work.
- •Configured MCP servers consume context even if not used
- •Too many servers reduce available context for reasoning and code
- •Use `/mcp` to audit connections and disable unused servers
When to prefer CLIs over MCP for context efficiency
Suggests using CLI equivalents when available because they don’t inject persistent tool definitions into context. Gives examples like GitHub (gh) and AWS (aws) tooling.
- •CLI tools can be more context-efficient than MCP servers
- •MCP adds persistent tool definitions; CLI use does not
- •Examples: `gh` for GitHub, `aws` for AWS
Using Skills and tool search mode to manage tool overload
Introduces “skills” as lightweight, on-demand context loaders (name + description) that can defer heavier details until needed. Explains that if MCP tools exceed ~10% of the context window, Claude Code may switch to tool search mode, which finds tools on demand but may be less reliable than having them already in context.
- •Skills load minimal metadata until Claude decides to use them
- •Skills can wrap CLI workflows as on-demand capabilities
- •If MCP tools exceed 10% of context, Claude Code may enter tool search mode
- •Tool search discovers tools on demand but may work less well than in-context tools
Final recap: how to add, scope, and keep MCP lean
Summarizes MCP’s purpose and the core operational advice: add servers, scope them for teams, and watch context usage. Reinforces the key commands and best practices for maintaining an efficient setup.
- •MCP connects Claude Code to external tools and data sources
- •Add servers with `claude mcp add`
- •Use `.mcp.json` project scope to share setup across a team
- •Monitor context usage and disable servers you don’t need
