Model Context Protocol (MCP) Guides

What is MCP (Model Context Protocol)?

MCP — the Model Context Protocol — is an open standard, originally introduced by Anthropic in late 2024, for connecting AI applications (like Claude Code, Claude Desktop, Cursor, and others) to external systems (databases, APIs, cloud providers, dev tools) in a uniform, reusable way. Think of it as USB-C for AI tools: one plug shape that lets any compatible LLM talk to any compatible service.

Before MCP, every AI application had to invent its own way to expose your database, your GitHub, your Slack, or your files. Each one had a bespoke integration, a bespoke auth flow, and a bespoke schema for tool inputs. MCP replaces that mess with a shared protocol: an MCP server (which exposes a specific system, e.g., Postgres) can be consumed by any MCP client (Claude Code, Claude Desktop, Cursor, Cline, and dozens more).

The MCP architecture in three parts

Every MCP interaction has three pieces:

• MCP Server — a program that exposes a specific system: Postgres, GitHub, Stripe, Slack, filesystem, etc. Runs as a subprocess (stdio) or remote HTTP endpoint.
• MCP Client — the AI app that connects to servers. Claude Code is a client. So is Claude Desktop, Cursor, Cline, Continue, Windsurf, and many others.
• Transport — how they talk. Two options: stdio (local subprocess, JSON-RPC over stdin/stdout) or http (streamable HTTP with SSE for remote servers).

The client discovers what the server can do, and the LLM decides which tools/resources to use based on the ongoing conversation.

What an MCP server can expose

• Tools — actions the LLM can invoke. execute_sql, create_github_issue, send_slack_message. Tools have JSON schemas for inputs and typed outputs.
• Resources — readable data the LLM can access. A file, a table's schema, a config document. Resources have URIs and can be watched for updates.
• Prompts — reusable prompt templates the user can invoke. Similar to slash commands but coming from the server.

Adding an MCP server to Claude Code

Two ways: the interactive command, or direct settings.json editing. Both produce the same result.

Or edit .claude/settings.json directly for team sharing:

Commit that file to git, and every teammate gets the same MCP setup on next pull. Environment variables like ${GH_TOKEN} stay in each developer's shell, so no secrets are committed.

Why MCP is a big deal (and why every serious Claude Code user runs 5+ MCP servers)

• One integration, many clients. The Postgres MCP server works identically in Claude Code, Claude Desktop, Cursor, Cline, Windsurf, and any other MCP client. Build once, use everywhere.
• Community ecosystem. There are hundreds of open-source MCP servers — databases, SaaS APIs, cloud providers, design tools, communication platforms. The catalog grows every week.
• First-party servers from major vendors. GitHub, Linear, Notion, Sentry, Cloudflare, Anthropic, and others ship their own MCP servers. Not community wrappers — official.
• Separation of concerns. The MCP server handles auth, rate limits, and API quirks. Claude just calls tools. Your codebase stays clean of API integration cruft.
• Remote or local. stdio transport for local process servers. http (streamable HTTP + SSE) for remote hosted servers — useful when the server needs cloud infrastructure (e.g., LLM-heavy servers).
• Standard auth patterns. OAuth 2.1 and PAT patterns are standardized. Your Linear MCP server sets up the same way as your Notion one.

MCP tools vs Claude Code's built-in tools

Claude Code has built-in tools (Read, Write, Edit, Bash, Grep, Glob) that operate on your local filesystem. MCP tools extend that reach to any system your MCP servers connect to:

• Read your Postgres schema? MCP.
• Create a GitHub issue? MCP.
• Look up a Linear ticket? MCP.
• Check Sentry for the latest error? MCP.
• Read a Notion page? MCP.
• Deploy to Cloudflare Workers? MCP.

Every MCP tool you add is one more system Claude can operate in. That's the whole game — giving Claude Code the same reach as a human developer with all their tabs open.

MCP vs hooks vs subagents

These three shape Claude Code in different ways:

• MCP = extending reach. Adds tools/resources that let Claude operate on external systems. Use when Claude needs to talk to something outside your filesystem.
• Hooks = enforcing policy. Runs shell commands at lifecycle events to gate, log, or observe. Use for security and standards. See hooks hub.
• Subagents = specialized delegation. Purpose-built configs with their own context. Use for role-based work. See agents hub.

They compose beautifully: a security-reviewer subagent reads your database schema via a Postgres MCP server, with all its tool calls audit-logged by a PostToolUse hook.