MCP Servers | AI Engineering Curriculum

What MCP Is

Model Context Protocol is an open standard created by Anthropic that lets Claude connect to external tools, data sources, and APIs. Think of it as a universal plugin system for AI.

Without MCP, Claude only works with what's in its context window. With MCP, it can talk to GitHub, query your database, search Slack, control a browser, call your internal APIs, and hundreds more.

The mental model: MCP servers expose tools Claude can call, just like its built-in tools (Read, Write, Bash). The difference is these tools come from the outside world.

What You Can Do With MCP

"Add the feature from JIRA issue ENG-4521 and create a PR on GitHub."
"Check Sentry for errors introduced by the last deployment."
"Find users who used this feature in our PostgreSQL database."
"Update the email template based on the new Figma designs in Slack."
"Create Gmail drafts inviting those users to a feedback session."

All in one session, driven by natural language.

Three Server Types

Type	How it runs	Use when
HTTP	Remote server	Cloud services - recommended
SSE	Remote server, older	Legacy - use HTTP instead
stdio	Local process	Local tools, custom scripts

Installing MCP Servers

HTTP (cloud services):

Bash

claude mcp add --transport http <name> <url>

# Examples
claude mcp add --transport http github https://api.githubcopilot.com/mcp/
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# With auth header
claude mcp add --transport http my-api https://api.example.com/mcp \
  --header "Authorization: Bearer your-token"

stdio (local tools):

Bash

claude mcp add --transport stdio <name> -- <command> [args...]

# Example: PostgreSQL database
claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
  --dsn "postgresql://user:pass@localhost:5432/mydb"

# Example: with env vars
claude mcp add --transport stdio airtable \
  --env AIRTABLE_API_KEY=your-key \
  -- npx -y airtable-mcp-server

Windows: Wrap npx with cmd /c: -- cmd /c npx -y @some/package

The Three Scopes

Scope	Stored in	Visible to
`local` (default)	`~/.claude.json`	Just you, current project
`project`	`.mcp.json` in project root	Your whole team (commit this file)
`user`	`~/.claude.json`	Just you, all projects

Bash

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --scope project
claude mcp add --transport http notion https://mcp.notion.com/mcp --scope user

When same server name exists at multiple scopes: local > project > user.

Managing Servers

Bash

claude mcp list          # list all servers
claude mcp get github    # details on a specific server
claude mcp remove github # remove a server
/mcp                     # check status inside Claude Code

Authentication

Bash

# 1. Add the server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# 2. Inside Claude Code:
/mcp
# Select server → Authenticate → follow browser OAuth flow
# Tokens stored securely and auto-refreshed

Popular MCP Servers

Server	What it adds
GitHub	Issues, PRs, repos, code review
Sentry	Error tracking, stack traces
Notion	Read/write pages and databases
Slack	Read/send messages
PostgreSQL	Query your database
Filesystem	Extended file system access
Airtable	Read/write Airtable bases

Find hundreds more: github.com/modelcontextprotocol/servers

MCP Resources - The @ Mention System

MCP servers expose resources you can reference with @ like files:

> Analyze @github:issue://456 and suggest a fix
> Compare @postgres:schema://users with @docs:file://database/user-model

Type @ in Claude Code to see all available resources in autocomplete.

MCP Prompts as Slash Commands

Servers can expose prompts as / commands:

/mcp__github__list_prs
/mcp__github__pr_review 456
/mcp__jira__create_issue "Bug in login flow" high

Format: /mcp__<servername>__<promptname> [arguments]

Using Claude Code as an MCP Server

Run Claude Code itself as an MCP server to expose its tools to other apps:

Bash

claude mcp serve

Add to Claude Desktop config:

JSON

{
  "mcpServers": {
    "claude-code": {
      "type": "stdio",
      "command": "claude",
      "args": ["mcp", "serve"]
    }
  }
}

Tool Search

When you have many servers, their tool definitions consume context. Tool Search loads tools on-demand instead:

Bash

ENABLE_TOOL_SEARCH=auto    # default - activates at 10% of context
ENABLE_TOOL_SEARCH=auto:5  # activate at 5% threshold
ENABLE_TOOL_SEARCH=true    # always on
ENABLE_TOOL_SEARCH=false   # always off

Environment Variables in `.mcp.json`

Share configs with your team without exposing credentials:

JSON

{
  "mcpServers": {
    "api-server": {
      "type": "http",
      "url": "${API_BASE_URL:-https://api.example.com}/mcp",
      "headers": {
        "Authorization": "Bearer ${API_KEY}"
      }
    }
  }
}

${VAR} - required | ${VAR:-default} - with fallback

Security Notes

Only install MCP servers you trust - they run with your permissions
stdio servers can access your file system
Servers fetching untrusted web content carry prompt injection risk
For teams: managed-mcp.json lets admins lock down which servers employees can use

Sources

Claude Code MCP Documentation (official docs)
Model Context Protocol (open standard)

What MCP Is

What You Can Do With MCP

Three Server Types

Installing MCP Servers

The Three Scopes

Managing Servers

Authentication

Popular MCP Servers

MCP Resources - The @ Mention System

MCP Prompts as Slash Commands

Using Claude Code as an MCP Server

Tool Search

Environment Variables in .mcp.json

Security Notes

Sources

Environment Variables in `.mcp.json`