07: MCP & External Tools

35 minutes | You need: GitHub access (or another external tool)

How MCP Works

Without MCP, Claude can read files and run shell commands — that’s it. MCP (Model Context Protocol) adds structured access to external systems: GitHub, databases, browsers, monitoring tools.

But MCP has a cost. Every MCP server exposes tools, and each tool has a description that consumes context tokens. Before you connect anything, understand the tradeoff.

MCP vs Skill+CLI: the decision framework

If the tool has a good CLI, prefer a skill that wraps the CLI:

Approach	Idle context cost	Good for
Skill wrapping CLI	~100 tokens (name + description)	`gh`, `aws`, `docker`, `kubectl`, `terraform`, `gcloud`
MCP server	Hundreds-thousands of tokens	No CLI exists, persistent connections (databases), structured I/O

Use MCP when:

No CLI exists for the tool
Persistent connection needed (database queries, websockets)
Structured input/output that CLI doesn’t handle well (complex API interactions)

Don’t install an MCP server for something gh or aws cli already does well. A skill that wraps the CLI loads only when invoked and costs nothing at idle.

Do This

1. Connect GitHub MCP

GitHub is the most common first connection:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

Verify:

/mcp

You should see GitHub tools in the list.

2. Do real work with it

Don’t just list PRs — do something useful:

Review PR #[number]. Check for: security issues, missing error handling, breaking changes to public APIs. Summarize findings with severity levels.

Or combine GitHub context with codebase work:

Check the 3 most recent open issues. For the one most related to authentication, look at the relevant code and draft a fix.

3. Connect one more tool

Choose based on your stack:

# PostgreSQL — natural language database queries
claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres

# Playwright — browser automation and E2E testing
claude mcp add playwright -- npx -y @anthropic-ai/mcp-server-playwright

# Sentry — production error analysis
claude mcp add sentry -e SENTRY_AUTH_TOKEN=your-token -- npx -y @sentry/mcp-server

# Notion — documentation and knowledge base
claude mcp add notion -e NOTION_API_TOKEN=your-token -- npx -y @makenotion/notion-mcp-server

Verify with /mcp again. Run /context to see the token impact of the new tools.

4. Understand Tool Search

On Sonnet 4+ and Opus 4+, Claude doesn’t load all tool descriptions into context upfront. Instead, Tool Search lazy-loads descriptions only when they seem relevant to the current task.

This reduces context bloat but isn’t free — each search query costs tokens. The practical impact:

You can connect many MCP servers without immediate context explosion
You still pay a token cost when Claude searches for relevant tools
Fewer servers = less noise in tool search results = better tool selection

Start with 2-3 MCP servers that match your immediate workflow. Add more only when you have a concrete need.

5. Scope and security

MCP configs live in three places:

Scope	File	When to use
Global	`~/.claude.json`	Personal tools across all projects
Project (shared)	`.mcp.json` in project root	Team tools — commit to git
Project (local)	`.claude/settings.local.json`	Personal credentials — gitignored

For team setups: Put shared server configs (without tokens) in .mcp.json. Keep tokens and personal credentials in .claude/settings.local.json which is gitignored.

Credentials: Use the -e flag to pass environment variables to MCP servers:

claude mcp add sentry -e SENTRY_AUTH_TOKEN=your-token -- npx -y @sentry/mcp-server

Never put tokens directly in .mcp.json if it’s committed to git.

6. Combine MCP with skills

The most powerful pattern is an MCP server providing data access, wrapped in a skill that provides workflow instructions:

---
name: triage-issues
description: Triage open GitHub issues by priority and suggest which to work on next
context: fork
---

Use the GitHub MCP tools to:
1. List open issues labeled "bug" or "feature"
2. For each, check: how old, how many comments, is it blocking anything
3. Cross-reference with the codebase — which issues touch areas with recent churn
4. Rank by: severity × recency × code-area-risk
5. Present top 3 recommendations with reasoning

The skill runs in a subagent (context: fork), uses MCP tools for GitHub access, and doesn’t pollute your main context.

Artifact

Two working MCP connections configured and tested with real tasks. Understanding of the MCP vs skill+CLI tradeoff.

Go Deeper

Playbook M05 — Agents and MCP for agent architecture and protocol details. Playbook M06 — Tool Design for building custom MCP servers for your internal tools.