Recommended MCP Servers for Claude | Official, Popular, and How to Choose by Use Case

Why Your "Selection Criteria" for MCP Servers Actually Matter

MCP (Model Context Protocol) is a standard protocol for connecting AI to external systems — Anthropic has described it as "the USB-C of AI" in their official blogsource. It works the same way across Claude.ai, Claude Code, and Claude Cowork, so users only need to decide which server to add; the client handles the rest.

That said, as of 2026 there are hundreds of MCP servers available — both official and third-party. Adding them without a clear purpose not only wastes your limited connection slots, but also risks granting the AI unnecessary permissions. The Free plan allows just 1 connection, and even Pro users have a limited practical quota, so your initial selection has a big impact on the overall experience.

Selection comes down to four criteria:

1. Official reference implementation or community-made? (trustworthiness)
2. Runs locally or connects remotely? (security)
3. Self-contained processing or does it require an external SaaS connection? (scope)
4. Maintenance status (recent commits, issue response time)

Keep these four points in mind as you read through the MCP servers below.

7 Official Reference MCP Servers

The official modelcontextprotocol/servers repository includes 7 MCP servers published as reference implementationssource. They are positioned as educational and reference implementations, but they are fully practical for local use and make an ideal starting point.

Filesystem and Git in particular are directly tied to local development productivity — they're the go-to first installs for most Claude Code users. Fetch is useful for passing external web content to Claude, and Memory enables knowledge retention across conversations. Sequential Thinking is well-regarded for structuring Claude's reasoning on complex tasks, with multiple use cases highlighted in Anthropic's official blog.

Note that GitHub, GitLab, Google Drive, Slack, PostgreSQL, and Puppeteer — which were previously included in this repository — have since been moved to the servers-archived repository. The recommended path is to use the third-party implementations described below as replacements.

Popular Third-Party MCP Servers

SaaS integrations and production-grade implementations not covered by the official references are provided by various companies and the community. Here they are in order of popularity among Claude users.

GitHub MCP

An officially maintained MCP server from GitHub that lets you perform repository operations, manage Issues and PRs, and run code reviews directly from Claude. Combined with Claude Code, automating code reviews and formatting PR comments is a typical use case. Connect via OAuth and restrict the scope to only the repositories you need for safety.

Notion MCP

Notion's official MCP server supports database read/write, page creation, and search. It's ideal when you want to handle internal knowledge searches entirely within Claude.ai, and connects via Connectors.

Slack MCP

Multiple third-party implementations exist, supporting message retrieval, posting, and channel management. Having Claude summarize and categorize support Slack threads is an increasingly common workflow, making this MCP a natural fit for support operations.

PostgreSQL / Supabase MCP

Among database MCPs, PostgreSQL and Supabase are the most popular — they let Claude execute SQL directly. Supabase provides an official MCP server with an architecture that holds credentials on the MCP side. The standard practice is to connect using a read-only role.

Puppeteer / Playwright MCP

Browser automation MCPs that let Claude control screenshots, form input, and E2E tests. They're a go-to example for frontend regression checks.

Brave Search MCP

A server that wraps the Brave web search API, allowing Claude to retrieve up-to-date information. Since Fetch alone assumes you already know the URL, Brave Search complements it by handling the search phase.

Recommended MCP Combinations by Use Case

The right "first MCP to add" depends on your use case. Here's a breakdown for three common personas.

For Developers Who Want to Accelerate Code Development with Claude Code

1. Filesystem (official)
2. Git (official)
3. GitHub (third-party)
4. Sequential Thinking (official)

This 4-server setup covers "local editing + Git operations + remote repository management + structured reasoning." It's a commonly adopted configuration for Claude Code users — just run claude mcp add four times and you're done.

For Business Users Who Want to Streamline Internal Workflows with Claude.ai

1. Notion (SaaS)
2. Slack (SaaS)
3. Google Drive (third-party)
4. Memory (official)

This configuration covers the four pillars: documents, communication, file sharing, and memory retention. Since everything connects via OAuth, plan for about 15 minutes for initial setup.

For Researchers Focused on Data Analysis and Investigation

1. Fetch (official)
2. Brave Search (third-party)
3. PostgreSQL or Supabase (DB)
4. Sequential Thinking (official)

This setup lets Claude run a full pipeline — actively pulling in external information, storing it in a database, and drawing conclusions through structured reasoning — all on its own.

Setup Instructions for Claude Code and Claude.ai

There are two ways to set things up, depending on your client.

Claude Code (CLI)

Run the following commands from your terminal:

claude mcp add filesystem -- npx -y @modelcontextprotocol/server-filesystem /Users/yourname/dev
claude mcp add git -- npx -y @modelcontextprotocol/server-git

Note that the path /Users/yourname/dev at the end of the first line is macOS-style. On Windows, use something like C:\Users\yourname\dev; on Linux, use /home/yourname/dev. Replace it with the path to your own working directory that you want Claude to have access to — this path defines the read/write scope for the Filesystem MCP.

Three connection types are supported: stdio, SSE, and Streamable HTTP. stdio connects to a local process via standard input/output, while SSE and Streamable HTTP connect to remote servers via URL. Servers launched locally with npx (as shown above) communicate via standard I/O, so stdio is the standard approach for local execution. Settings are saved to ~/.config/claude/mcp.json and similar locationssource.

If you want to use GitHub with Claude Code: the previously documented claude mcp add github -- npx -y @modelcontextprotocol/server-github is an old reference implementation that has been moved to servers-archived (educational only) and is now deprecated. As mentioned in the "GitHub MCP" section above, the recommended approach today is to use the official GitHub remote MCP server via OAuth — either by registering it as a remote (HTTP) server in Claude Code, or by connecting via OAuth through Claude.ai's Connectors. The Filesystem and Git entries in the code block above are official reference implementations that have NOT been archived, so using them via npx is perfectly fine.

Claude.ai (Browser)

In the browser version, go to Settings > Connectors > Add custom connector and register the remote MCP server's URL. For SaaS tools with built-in OAuth flows, simply granting permission on the authorization screen completes the connection. The Free plan has a 1-connection limit, so prioritize the tool you use most.

Claude Desktop (macOS / Windows)

Add server definitions as JSON to claude_desktop_config.json. You can use the same MCP server implementations as Claude Code. Check the official documentation for the configuration file pathsource.

A Must-Check List Before Selecting Any MCP Server

Finally, here are 5 things to verify before adding a new MCP:

1. Is it from the official repository (modelcontextprotocol/servers) or a trusted publisher?
2. Has there been a commit within the last 90 days? (maintenance status)
3. Does it run locally or connect remotely? (who controls the data being sent)
4. Are the required OAuth scopes proportionate to your actual use case?
5. If multiple MCPs offer similar functionality, have you compared them by Issue count and Star count?

In particular, watch out for cases where "read-only access would suffice, but write permissions are being requested." Prioritize implementations that let you restrict the scope. Anthropic itself repeatedly recommends "minimizing tool permissions" in their blogsource.

Conclusion — Start with 4 Official Servers, Then Add SaaS Integrations

While hundreds of MCP servers exist, the four you should install first are the official reference implementations: Filesystem, Git, Memory, and Sequential Thinking. These cover local development and reasoning support. From there, add third-party options like GitHub, Notion, Slack, and Brave Search as your workflows demand — and Claude becomes an AI client optimized for exactly how you work.

As the official repository notes, reference implementations are educational in nature. For production use cases, we recommend either actively maintained third-party implementations or building your own MCP server in-house. Python and TypeScript SDKs are available, and a minimal implementation can be started in as few as 50 lines of code.