The Best MCP Servers for Claude: A Curated Guide
The Model Context Protocol (MCP) is the USB standard of the Claude ecosystem: an open protocol that lets any tool expose capabilities — query this database, drive this browser, search these docs — to any AI client that speaks it. Claude Code, the Claude apps, and most AI-native editors all do. Anthropic open-sourced it; the ecosystem now spans thousands of servers.
Which creates the actual problem: thousands of servers, most of them abandoned weekend projects. This guide organizes by category of capability — install one good server per category you need, and skip the rest. Server names churn; the categories are stable.
How installation works
In Claude Code, adding a server is one command:
# example: add a server (local stdio process) claude mcp add <name> -- <command to run the server> # example: add a remote server claude mcp add --transport http <name> <url> # see what's installed claude mcp list
In the Claude desktop and web apps, servers are added as "connectors" from settings — same protocol, friendlier packaging.
The categories that earn their keep
1. Version control: GitHub / GitLab
The official GitHub MCP server gives Claude real access to issues, PRs, reviews, and CI status — so "look at the failing checks on my PR and fix them" becomes one request instead of five copy-pastes. If your work lives in GitHub, this is the first install.
2. Browser automation: Playwright-based servers
A browser server lets Claude open pages, click, fill forms, and read the DOM — which means it can verify its own frontend work end-to-end instead of declaring victory from the code. Microsoft's Playwright MCP server is the reference choice. For web developers this is arguably the highest-value connector on the page.
3. Databases: Postgres and friends
Database servers let Claude inspect schemas and run queries — real ones, against your dev or read-replica instance. Debugging "why is this query slow" or "what does this data actually look like" stops being a copy-paste relay. Safety rule: point it at development or read-only credentials, never production write access.
4. Documentation and search
Two flavors: web search servers (current information from the live web) and documentation servers like Context7 that feed Claude version-accurate library docs — which noticeably cuts "that API doesn't exist anymore" errors in generated code.
5. Errors and observability: Sentry et al.
Connect your error tracker and "investigate this Sentry issue" becomes a complete workflow: Claude reads the stack trace and breadcrumbs, finds the offending code, and proposes the fix with the context of how often and for whom it breaks.
6. Project and knowledge tools: Linear, Jira, Notion, Slack
Ticket context without tab-switching: "implement LIN-482" pulls the issue, acceptance criteria, and discussion. Valuable for teams whose specs actually live in these tools; skippable for solo work.
7. Design: Figma
Design-context servers let Claude read component structure and styles from Figma files — the difference between "build something like this screenshot" and "build this design, with its actual spacing and tokens."
The security section (read this one)
An MCP server is code running with real credentials. Three rules keep the power without the incident report:
- Least privilege: read-only tokens where possible, dev databases over prod, scoped API keys over account-wide ones.
- Provenance: prefer official vendor servers and widely-used open-source ones you can read. A random server from a listing site gets your GitHub token — act accordingly.
- Prompt injection awareness: anything Claude reads through a connector (web pages, tickets, docs) can contain text designed to manipulate it. Keep confirmation prompts on for consequential actions; don't run auto-approve against untrusted content.
A sane starter set
- GitHub — if your code lives there.
- Playwright — if you build anything with a UI.
- Your database — read-only, dev instance.
- One docs/search server — if you work across many libraries.
Then stop until a real task wants more. Every server adds tools to Claude's context — a lean set it uses well beats an arsenal it fumbles.
MCP gives Claude hands; skills give it expertise. The ClaudeThings kits ship 103 skills and 181 commands that put both to work — code review, security audits, SEO, email sequences — installed in one command. See what's inside →