claude-code-slides

MCP Servers — Setup Guide

Related slide: Slide 9 — Connect to Everything

Official docs: Claude Code MCP Setup

What is MCP?

Model Context Protocol (MCP) is an open standard that lets coding agents call external tools at runtime. Claude Code, OpenCode, and Gemini CLI all support it.

MCP servers give your agent access to databases, APIs, browsers, documentation, and more.

SAP MCP Registry

SAP maintains its own MCP registry with pre-approved servers:

https://ai-docs.portal.hyperspace.tools.sap/mcp-registry/

Check here first for SAP-specific integrations.

Adding MCP Servers

HTTP (remote servers — recommended)

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

# Example: Context7 for documentation
claude mcp add --transport http context7 https://mcp.context7.com/mcp

Stdio (local servers)

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

# Example: Playwright for browser automation
claude mcp add --transport stdio playwright -- npx -y @playwright/mcp@latest

Scopes

Scope Flag Stored in Use case
Local --scope local (default) ~/.claude.json Personal, one project
Project --scope project .mcp.json (version controlled) Shared with team
User --scope user ~/.claude.json Personal, all projects

Managing servers

claude mcp list              # List all configured servers
claude mcp get <name>        # Get details for a server
claude mcp remove <name>     # Remove a server
/mcp                         # Check server status (inside Claude Code)

My Common MCP Servers

Context7 — Documentation lookup

Fetches up-to-date library docs so the agent doesn’t hallucinate APIs.

claude mcp add --transport http context7 https://mcp.context7.com/mcp

When to use: When working with libraries where the agent’s training data may be outdated.

Chrome DevTools

Direct browser inspection via DevTools protocol. Useful for debugging, but consider the skill-based alternative below.

Prefer instead: The agent-browser skill — it’s a CLI tool that doesn’t require an MCP server, uses less context, and is more secure.

Playwright

Browser automation for testing and interaction.

claude mcp add --transport stdio playwright -- npx -y @playwright/mcp@latest

Prefer instead: playwright-cli — a CLI alternative that works as a skill rather than MCP, giving you deterministic control without the MCP overhead.

MCP vs CLI Skills — When to Use What

For most MCP servers, there’s a CLI-based skill alternative. Prefer CLI/skills when possible.

  MCP Server CLI Skill
Security Runs with your permissions, can be a vector for prompt injection Deterministic, code-based — no LLM interpretation
Context cost Tool definitions consume tokens even when idle Loaded only when invoked
Control Agent decides when/how to call tools You control invocation through skill triggers
Best for Services that need persistent connections (databases, APIs with auth) Tools that run a command and return output

Examples of CLI replacements

MCP Server CLI Alternative
Playwright MCP playwright-cli
Chrome DevTools MCP agent-browser skill
GitHub MCP gh CLI (already available)
AWS MCP aws CLI
Google Cloud MCP gcloud CLI

Security Warning

Only install MCP servers you trust.

A malicious MCP can read files, exfiltrate data, or execute arbitrary code — all with your permissions.

See the unsafe MCP demo for a practical example of how MCP servers can be exploited.

Protect yourself

  1. Only install from trusted sources — SAP registry, official vendors, or servers you’ve reviewed
  2. Use deny rules for sensitive files (see Security slide)
  3. Use hooks for deterministic enforcement (see Permissions slide)
  4. Prefer CLI tools over MCP when a CLI alternative exists
  5. Review project .mcp.json files before accepting — Claude Code prompts for approval, don’t auto-accept

Further Reading