Claude MCP Servers: Complete List and Setup Guide (2026)

A developer installed 15 MCP servers the day they set up Claude Code. Three months later, only five were still in their config. The rest? Forgotten, unused, and quietly eating context tokens in every single conversation.

That's the pattern across the Claude MCP community right now. There are hundreds of servers available. The hard part isn't finding them. It's knowing which ones actually improve your workflow, how to configure them without breaking things, and when to stop adding more.

Claude MCP servers extend what Claude can do beyond conversation. Hook up a GitHub server and Claude creates issues, reviews PRs, and searches code across your repos. Add a database server and it queries production data directly. Connect a browser server and it navigates real websites, even ones with Cloudflare protection. The Model Context Protocol makes all of this work across Claude Desktop, Claude Code, and Claude Web.

This guide covers every major server worth connecting, the setup configs you'll need, the problems the community keeps hitting, and the strategies that separate productive MCP users from frustrated ones.

How Claude MCP Works

The Model Context Protocol creates a standard way for AI clients like Claude to discover and call external tools. Each MCP server exposes tools through a JSON-RPC interface. When you connect a server to Claude, those tools become available in your conversation or coding session.

There are three ways to connect claude mcp servers:

Stdio (local servers) runs the server as a local process. Claude desktop mcp configurations use this transport primarily. You define the server in claude_desktop_config.json and it starts automatically when Claude launches.

HTTP/SSE (remote servers) runs the server on a remote URL. Claude code mcp setups and Claude Web both support this natively. You point them at the server's endpoint and they connect over HTTP with Server-Sent Events for streaming responses.

Gateway (managed) aggregates multiple servers behind a single endpoint. An MCP gateway handles authentication, routing, and tool filtering. You configure one connection instead of ten.

Each approach has tradeoffs. Stdio works well for solo local development but falls apart when your team needs shared access. Remote HTTP scales better but requires you to host the server somewhere. A gateway gives you both plus observability, at the cost of an extra infrastructure layer.

Complete List of Claude MCP Servers by Category

Here's every major MCP server that works with Claude, organized by what developers actually keep installed. Each links to its verified listing on the Apigene MCP directory where available.

Developer Tools

These are the servers working developers keep running after the initial excitement fades.

A highly-upvoted community tip worth noting: "Claude Code has built-in ability to use the gh CLI for GitHub, no MCP needed." If you're already in Claude Code, test whether the CLI covers your needs before adding the GitHub MCP server. The MCP server adds value when you want structured tool calls with typed responses rather than parsing shell output.

Web Browsing and Research

Claude's built-in web fetch gets blocked by Cloudflare and bot protection on most real websites. One developer tested it against 10 popular sites and it failed on all 10. These servers fill that gap.

Chrome DevTools MCP deserves a callout. It's "zero-configuration" from the Chrome team, and multiple developers described it as a breakthrough compared to typical MCP browser setup. If you don't need full automation (form filling, multi-page flows), it's the fastest path to giving Claude browser access.

Productivity and Communication

These connect Claude to the tools your team already uses.

Two warnings from the community. First, Slack MCP can be "too noisy" because the "agent kept wanting to post updates." Configure tool-level permissions so Claude reads channels but doesn't send messages without your explicit approval. Second, the Atlassian MCP "is so crap" according to one popular thread with 140 upvotes. Several developers recommend using direct Jira API calls through Claude Code's shell access instead.

Data and Analytics

One developer called DeepWiki "magical in Claude Code, especially for bleeding edge libraries or when documentation sucks." It's particularly useful when you're working with libraries too new for Claude's training data.

How to Set Up Claude MCP Servers

Claude Desktop MCP Configuration

Claude Desktop reads server configs from a JSON file. On macOS, it lives at ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows, it's historically been at %APPDATA%\Claude\, though updates have moved this path more than once (more on that in the troubleshooting section).

Here's a working config with two servers:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "ghp_your_token_here"
      }
    },
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-brave-search"],
      "env": {
        "BRAVE_API_KEY": "your_key_here"
      }
    }
  }
}

Save the file, restart Claude Desktop, and both servers will be available. You can verify they loaded in Claude Desktop's settings panel under "Local MCP Servers."

Claude Code MCP Configuration

Claude code mcp supports both local and remote servers. Add them directly from the terminal:

# Add a remote HTTP server
claude mcp add --transport http github https://api.githubcopilot.com/mcp/
 
# Add a local stdio server
claude mcp add --transport stdio brave-search npx -y @modelcontextprotocol/server-brave-search

Or configure them in your project's .mcp.json file so the whole team gets consistent server access:

{
  "mcpServers": {
    "postgres": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-postgres"],
      "env": {
        "DATABASE_URL": "postgresql://user:pass@host:5432/db"
      }
    }
  }
}

As of Claude Code 2.1.80, you can also configure MCP plugins inline via settings.json, which means fewer scattered config files across your project.

Claude Web (Remote MCP)

Claude Web supports remote MCP servers through OAuth-authenticated connections. This is where the setup gets tricky. The OAuth flow requires your server to implement .well-known discovery endpoints, and developers report that the process is "a bit tricky, and the SDK wasn't always the best help."

If you're hitting auth issues with remote claude mcp servers, check two things. First, make sure Cloudflare's "Block AI bots" setting isn't intercepting Claude's OAuth callbacks. Multiple developers spent hours debugging before discovering their WAF was the culprit. Second, note that Claude's OAuth client metadata endpoint moved from claude.ai/oauth/mcp-oauth-client-metadata to claude.ai/api/oauth/mcp-oauth-client-metadata in early 2026, breaking servers that hardcoded the old path.

The Token Problem: Why Fewer Servers Is Better

Every MCP server you connect adds its tool definitions to Claude's context window. This cost is invisible but significant. One developer measured the impact:

A separate benchmark found even worse numbers: 84 tools across several MCP servers consumed 15,540 tokens at session start, before the agent processed a single user message.

"Most people think they need a higher Claude plan when their usage runs out in 2 hours," one developer observed. "They don't. They need fewer MCP tools loaded." Another found that 30-40% of their context window went to tool definitions "that were never used, every time."

The community's consensus: start with 3-5 servers covering your daily workflow. Add more only when you hit a specific need. If you find yourself connecting 8+ servers, use an MCP gateway that provides selective tool loading, exposing only the tools relevant to your current task.

Common Setup Issues and Fixes

Based on 14 Reddit threads, these are the issues developers hit most often with claude mcp servers.

Windows Config Path Changes

Claude Desktop on Windows has changed its config file location multiple times. The MSIX packaging moved the expected path, and one developer called it "a gigantic mess" that was "really really bad." If your servers stopped working after an update, check both %APPDATA%\Claude\ and %LOCALAPPDATA%\Packages\Claude_\LocalCache\Roaming\ for your config file. Don't use symlinks between the old and new paths. They cause blank screen issues on restart.

Server Shows Connected But Tools Don't Work

The second most common issue: a server appears connected in Claude Desktop, but tool calls fail silently. For GitHub MCP specifically, this usually means the PAT isn't being passed correctly. Verify your token works with curl before assuming the server is broken. Also check for transport mismatches: HTTP-based MCP URLs "only work with VS Code/Cursor, not Claude Desktop" because Desktop expects stdio servers.

First Tool Call Fails, Retry Works

If you're running servers behind a reverse proxy like NGINX, the first tool discovery call often hangs. NGINX's default proxy_read_timeout of 60 seconds and its habit of buffering SSE streams cause silent failures. Increase the timeout and add proxy_buffering off for your MCP endpoints.

Tool Name Collisions

When you have multiple MCP servers configured, conflicting tool names across servers cause failures "with no useful message." One developer spent an hour toggling connectors before discovering an underscore-versus-hyphen mismatch in a tool name. If tools suddenly show as "disabled in your connector settings" despite being enabled, check for name conflicts between servers.

Tool-Use Limit Interruptions

A newer issue as of March 2026: Claude Desktop started stopping mid-task with "Claude reached its tool-use limit for this turn." One developer measured the same MCP workflow going from 6% to 14% of the 5-hour usage limit. If you're hitting this, reducing the number of loaded tool definitions can help because fewer tools means fewer implicit tool-call decisions per turn.

Best Claude MCP Servers: What Developers Actually Keep

After analyzing what stays installed after months of daily use (not what gets installed on day one), here are the best claude mcp servers by real retention:

1. Filesystem + Git (built-in) - The defaults. Developers who tried third-party file management MCPs came back to these.
2. GitHub MCP - Create issues, review PRs, search code. Replaces having GitHub open in a browser tab.
3. Postgres/Supabase - Query databases directly from conversation. Eliminates context-switching to a SQL client.
4. Playwright or Chrome DevTools - Browser access that actually works on modern websites with bot protection.
5. Brave Search or Exa - Web search that gives Claude real internet access beyond its training data.
6. Context7 - Library documentation that's always up to date, even for brand-new packages.
7. Sentry - Pull error traces directly into your debugging context.

The pattern: developers keep servers that replace a browser tab they'd otherwise need open. Servers that duplicate functionality Claude already has through CLI access or built-in tools get removed.

When to Use a Gateway Instead of Direct Connections

Direct MCP connections work for solo developers with 3-5 servers. Once you're in any of these situations, a gateway becomes the practical choice:

• Team access: Multiple people need the same servers with consistent configuration
• Too many tools: Your context window fills up with unused tool definitions
• Auth management: You're storing API keys in config files across multiple machines
• Observability: You need to know which tools get called, how often, and what fails
• Security: You want per-tool permissions so agents can read but not write without approval

Apigene's MCP gateway aggregates 251+ vendor-verified servers behind a single endpoint. You configure one connection in Claude, and the gateway handles auth, tool filtering, and routing. It also provides the observability that direct stdio connections can't: per-call logs, token usage tracking, and error reporting across every server.

For developers just getting started with claude mcp servers, direct connections are the right call. But if you're building anything production-grade or shared across a team, the gateway saves you from the configuration drift and token overhead that frustrate experienced MCP users.