Claude: Zero to Hero
Claude Code · module 9 of 13

MCP in Claude Code

What you’ll learn: how to connect Claude Code to your database, ticket system, browser and internal tools — and how to keep the context cost near zero.

Recap

The Model Context Protocol is an open standard. An MCP server exposes tools; Claude Code is the client that discovers and calls them.

Without MCP, Claude Code can read your repo and run shell commands. With MCP, it can query your staging database, read a Jira ticket, check a Sentry error, or drive a browser — as first-class tools, not shell hacks.

Adding a server

# Fastest path
claude mcp add <name> <command> [args...]

# Or from a config file
claude --mcp-config ./mcp.json

# Only use servers from that file, ignoring everything else
claude --strict-mcp-config --mcp-config ./mcp.json

Config shape:

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres", "postgresql://localhost/mydb"]
    },
    "sentry": {
      "type": "http",
      "url": "https://mcp.sentry.dev/mcp"
    }
  }
}

Transports: stdio (a local process), HTTP, and SSE.

Scope

Scope Where Shared
Local Your machine, this project No
Project .mcp.json in the repo Yes, via git
User ~/.claude.json You, all projects

Precedence: local > project > user.

Project scope is the good one for teams — commit .mcp.json and everyone gets the same tools.

Authentication

claude mcp login sentry           # run the OAuth flow from the shell
claude mcp login sentry --no-browser   # over SSH: prints the URL to paste
claude mcp logout sentry

Or use /mcp inside a session for the interactive panel.

Context cost, and why it’s low

Tool search is on by default. At session start, Claude Code loads only tool names from connected servers; full JSON schemas stay deferred until Claude needs a specific tool.

The consequence: idle MCP servers cost very little. You can connect several without paying for all of them on every request.

Check the actual cost:

/mcp

Shows connection status and token cost per server. Disconnect what you’re not using.

Claude Code reconnects to remote servers automatically if they drop.

Servers worth connecting

Server Why
Database (Postgres, MySQL, SQLite) Claude can inspect the real schema and run real queries instead of guessing
Playwright / browser Verification. Claude can actually check the UI works, which matters enormously for autonomous work
Sentry / error tracking Go from an error to a fix without leaving the terminal
GitHub / GitLab Issues, PRs, reviews
Jira / Linear Ticket context in the session
Filesystem (scoped) Read a directory outside the repo
Your own internal tools Usually the highest-value one

Browse: Connectors directory.

The Playwright/browser one deserves emphasis. Anthropic’s own guidance on long-horizon agentic work says: “As the length of autonomous tasks grows, Claude needs to verify correctness without continuous human feedback. Tools like Playwright MCP server or computer use capabilities for testing UIs are helpful.” Verification tooling is what makes long autonomous runs actually work.

Writing your own MCP server

You don’t need to unless you have a system worth exposing. When you do:

Tool design rules that matter:

  1. Descriptions are prompts. Claude picks tools by matching your description. Say what it does and when to use it.
  2. Few, well-scoped tools beat many overlapping ones. Three tools with clear boundaries outperform twelve with fuzzy ones.
  3. Return useful errors. “Query failed” teaches Claude nothing. “Column user_id does not exist on table orders; did you mean customer_id?” lets it self-correct.
  4. Keep results small. Tool results enter the context window. Paginate. Summarise. Don’t return 10,000 rows.
  5. Mark destructive tools. Set requiresUserInteraction on anything irreversible so it always prompts.

Security

MCP servers are the widest attack surface you can add to Claude Code.

A server can: read anything it has access to, take actions with its credentials, and return content that Claude reads as context — which means it can inject instructions.

Practical controls:

Treat connecting a third-party MCP server like adding a dependency with production credentials — because that’s what it is.

MCP tunnels

For enterprise cases where an MCP server lives inside your network and needs to be reachable by Anthropic-hosted surfaces, MCP tunnels provide the connection with certificate-based auth. Deployable via Docker Compose or Helm, manageable in the Console.

See MCP tunnels.

Channels

A related research-preview mechanism: an MCP server can push events into a running Claude Code session — CI results, chat messages, monitoring alerts — so Claude reacts while you’re away.

claude --channels plugin:my-notifier@my-marketplace

See Channels.

Try it

Exercise 1 — Database server. Connect a Postgres or SQLite MCP server to a dev database. Ask Claude a question requiring schema knowledge. Compare against what it guessed before.

Exercise 2 — Cost check. Run /mcp before and after connecting two servers. Note the token cost. Then note how it doesn’t change much — that’s tool search working.

Exercise 3 — Verification loop. Connect a browser MCP server. Ask Claude to make a UI change and verify it works. This is the difference between “Claude wrote code” and “Claude shipped a change.”

Exercise 4 — Project scope. Create .mcp.json at your repo root with the servers your team needs. Commit it. Onboarding just got shorter.

Exercise 5 — Gate an MCP tool. Add a PreToolUse hook with matcher mcp__.* that logs every MCP call. Read the log after a day.

Exercise 6 — Error quality. If you build a server: deliberately return a bad error message, watch Claude flounder, then improve it and watch it self-correct. This is the single biggest lever on tool usability.

Checkpoint

Going deeper