MCP Servers for Claude Code: What They Are and 10 Worth Installing

Most Claude Code users have either never heard of MCP or have heard of it, tried to set one up, got lost in JSON configs, and gave up. I have done both. This post is the version I wish someone had handed me.

What MCP is, in one paragraph

MCP stands for Model Context Protocol. It is an open protocol from Anthropic that lets Claude Code talk to external tools — databases, browsers, APIs, your filesystem, anything — through a standardized interface. You install an MCP server, Claude Code connects to it, and suddenly Claude has new tools it can call. That is the whole concept. Everything else is implementation detail.

The reason it matters: without MCP, Claude Code only knows what is in your codebase and what you tell it. With MCP, Claude Code can query your database, browse the web, read your Notion, check your Linear tickets, inspect a running server, or literally anything else someone has written a server for.

How to install one

All MCP setup in Claude Code goes through the claude mcp CLI. The common pattern:

claude mcp add <name> <command to run the server>

For example, the official filesystem server:

claude mcp add filesystem npx -y @modelcontextprotocol/server-filesystem /Users/you/projects

That adds a filesystem MCP server scoped to your projects directory. Claude Code can now read any file in that tree without needing you to pass each one.

There is also claude mcp list, claude mcp remove <name>, and a global settings file at ~/.claude/mcp.json if you want to edit by hand. But 95% of the time the CLI is faster.

The 10 MCP servers actually worth installing

1. Filesystem

The starter server. Gives Claude controlled access to a directory tree without per-file prompting. If you only install one MCP server, install this one.

claude mcp add fs npx -y @modelcontextprotocol/server-filesystem ~/projects

2. GitHub

Browse repos, read issues and PRs, search across your org, comment on threads. Dramatically better than asking Claude to "go look at that PR" without context.

claude mcp add github npx -y @modelcontextprotocol/server-github

Set GITHUB_PERSONAL_ACCESS_TOKEN in the environment before running.

3. Puppeteer / Playwright

Lets Claude drive a real headless browser. Screenshot a page, fill a form, click through a flow, capture network traffic. I use this constantly for frontend debugging — "take a screenshot, tell me why this button is misaligned."

claude mcp add playwright npx -y @playwright/mcp

4. Postgres

Query your local or staging database directly. Claude writes the SQL, runs it, interprets results. Much faster than copy-pasting query output back and forth.

claude mcp add postgres npx -y @modelcontextprotocol/server-postgres "postgresql://..."

Read-only connection strings only, please. Do not point production at this.

5. SQLite

Same idea as Postgres but for SQLite files. I use it for D1 development — point it at my local .wrangler/state/d1/*.sqlite and Claude can debug schema issues or seed data without me running commands.

6. Slack

Read recent messages, post to channels, DM yourself a summary. Tricky to set up (OAuth) but worth it for team workflows where the answer to "why is staging down" is sitting in a Slack thread somewhere.

7. Linear / Jira / ClickUp

Whatever your team uses for tickets. Claude can fetch the ticket you are working on, read the description, check the subtasks. No more pasting the whole ticket into the prompt.

8. Brave Search / Exa

Web search. The built-in WebSearch tool is fine but has rate limits and can be flaky. A dedicated MCP server is faster and returns more consistent results.

claude mcp add brave-search npx -y @modelcontextprotocol/server-brave-search

9. Sequential Thinking

This one is weird but useful. It is an MCP server that gives Claude a "scratchpad" tool for breaking down complex problems. Claude calls thought(1 of 5), thought(2 of 5), etc. and the tool responses keep it structured. Works especially well for architecture decisions.

10. Context7

Fetches up-to-date library documentation. Claude's training cutoff means it sometimes hallucinates API signatures for libraries that have moved on. Context7 grabs current docs for the exact version you are using.

claude mcp add context7 npx -y @upstash/context7-mcp

The honest downsides

MCP is not a silver bullet:

Every server you add increases Claude's tool count. Past a dozen tools, model performance on tool selection starts to degrade.
Some servers are buggy. Community servers especially — check the GitHub issues before trusting one with production access.
Debugging a broken MCP server is painful. Logs go to stderr, and if the handshake fails the server silently does not load.
Auth is per-server and mostly via environment variables. Keeping secrets tidy takes discipline.

My rule: install a server when I can name the workflow it unblocks. If I cannot think of three specific tasks I would use it for, I skip it.

What this has to do with MOLTamp

MOLTamp is where you run Claude Code. Claude Code is where you run MCP servers. The three layers stack.

A good MCP setup plus a well-tuned Claude Code plus a MOLTamp cockpit you actually enjoy looking at — that is the full AI development rig in 2026. Skip any one and you are leaving productivity or vibes on the table.

If you have not added a single MCP server yet, start with filesystem. Thirty seconds of setup, daily payoff.