MCP Servers and Plugins

MCP connects Claude Code to the systems your work actually lives in: tickets, errors, databases, browsers. Plugins package everything from the previous chapters so a team can install a whole workflow in one command. These are the two primitives that scale past what one engineer configures by hand.

MCP in one paragraph

The Model Context Protocol is an open standard for exposing tools and data to AI applications. A server implements the protocol once and works with any MCP client, including Claude Code. I have described it as doing for AI what USB did for hardware; the longer explanation is in MCP: A Comprehensive Guide.

Adding a server

# Remote server over HTTP (GitHub's hosted MCP server)
claude mcp add --transport http github https://api.githubcopilot.com/mcp/

# Another remote server
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

# Local server launched as a process (stdio)
claude mcp add playwright -- npx -y @playwright/mcp@latest

# Pass environment variables
claude mcp add db --env DATABASE_URL=postgres://readonly@... -- npx -y your-db-mcp

# Manage
claude mcp list
claude mcp remove playwright

Inside a session, /mcp shows connection status and handles OAuth sign-in for remote servers.

Servers can be scoped to you, to a project, or shared with the team through a committed .mcp.json at the repository root. Commit the configuration, never the credentials. Reference secrets through environment variables.

Servers worth running

Server type	What it gives Claude
GitHub or GitLab	Read issues, review PRs, check CI status, open PRs
Linear, Jira, or Asana	Pull ticket context, update status, link commits
Sentry or similar error tracker	Pull stack traces and frequency for a real root cause
Playwright or another browser server	Drive a real browser for UI checks and exploratory testing
Read-only database	Inspect schemas and real data shapes
Docs or search server	Fetch current library documentation
Slack	Read incident threads, post summaries

The workflow change is significant. "Fix the bug in ticket ENG-412" becomes a complete instruction: Claude reads the ticket, pulls the linked Sentry trace, finds the code, fixes it, and references the ticket in the commit.

MCP over raw CLI for sensitive systems

You could let Claude run psql with production credentials. Do not. An MCP server lets you expose exactly the operations you intend, such as read-only queries against specific schemas, with logging you control. Anthropic's own teams recommend MCP over raw CLI access for sensitive data for this reason.

Rules I apply:

• Database servers connect with a read-only role, ideally to a replica or sanitized copy.
• Production write access through MCP requires a human confirmation step.
• Every server's tool list is reviewed before it is added. Disable tools you do not need.

MCP risks to take seriously

MCP widens what Claude can reach, which widens what can go wrong.

• Prompt injection through data. A ticket, web page, or log line can contain text written to manipulate the agent. Treat tool output as untrusted input.
• Tool poisoning. A malicious or compromised server can describe its tools in misleading ways. Install servers only from sources you trust and pin versions.
• Excess privilege. Broad tokens turn a small mistake into a large incident. Scope every credential to the minimum.

These are identity and access problems at their core. I cover the agent side in depth in my writing on AI agents and identity.

Context cost of MCP

Every connected server adds tool descriptions to context. Ten servers with twenty tools each is a lot of always-present text. Claude Code now defers loading many tool schemas until needed, which helps, but the discipline still applies: connect what the project uses, disconnect what it does not.

Plugins

A plugin bundles skills, subagents, hooks, slash commands, and MCP server configuration into one installable package. A marketplace is a catalog of plugins, often a GitHub repository.

/plugin                                         # browse and manage
/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Plugins are how good setups spread. One engineer builds a reviewer subagent, a formatter hook, and three skills; packaged as a plugin, the whole team gets them in a minute and updates flow automatically.

Plugins worth starting with

• The official Anthropic marketplace, reachable from /plugin, including code review, security, and frontend design plugins.
• Superpowers, a disciplined workflow pack: brainstorm, write a plan, execute the plan in small verified steps, with strong TDD and debugging skills.
• Language and code-intelligence plugins for typed languages, which give Claude precise symbol navigation and error detection after edits.

Building an internal marketplace

For a team, create a private repository that acts as a marketplace. Each plugin lives in its own folder with a manifest. Engineers add the marketplace once and install the plugins they need. Changes go through pull requests like any other shared code, which gives you review and history for your AI tooling.

Key takeaways

• MCP connects Claude to the systems your work lives in, from tickets to error traces.
• Prefer scoped MCP servers with read-only roles over raw CLI access to sensitive systems.
• Treat tool output as untrusted input; prompt injection and tool poisoning are real.
• Plugins spread good setups across a team; review and pin them like any dependency.