Complete Guide to MCP Servers: What They Are and How to Use Them

Model Context Protocol (MCP) servers give AI coding agents access to external tools and data. Instead of pasting context by hand, your agent can read GitHub issues, query databases, search Notion, and control a browser. This guide explains what MCP is, how to set it up, and how to choose servers for your workflow.

Quick Answer

MCP is an open standard (by Anthropic) that lets AI applications talk to external systems through a uniform protocol. MCP servers are processes that expose tools (functions) and resources (data) to AI agents. You add servers to your AI tool's config; the agent discovers and calls them at runtime.

What is MCP?

MCP defines a JSON-RPC 2.0 protocol for AI applications to connect to external data and tools. A host (Cursor, Claude Code, Codex) runs an MCP client that talks to one or more MCP servers. Each server exposes:

• Tools — Callable functions (e.g., create_issue, execute_sql)
• Resources — Readable data (e.g., database schemas, file contents)
• Prompts — Reusable prompt templates

Two common transport modes:

• STDIO — The host spawns the server as a local process; communication over stdin/stdout.
• HTTP/SSE — The server runs remotely; the client connects via URL (often with OAuth).

Practical Setup Walkthrough

Cursor

1. Create .cursor/mcp.json in your project root (or ~/.cursor/mcp.json for global config).
2. Add a server. Example for GitHub:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_your_token"
      }
    }
  }
}

3. Restart Cursor. In Settings > Tools & MCP you should see the server listed.
4. In a chat, try: "Search my repo for uses of fetchUser."

Claude Code / Claude Desktop

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Use the same mcpServers structure. Restart Claude after editing.

VS Code with GitHub Copilot

Some clients support code --add-mcp '...' or a GUI for adding servers. Check your client's MCP documentation.

Remote servers (Supabase, Sentry, Vercel)

Many SaaS MCP servers use a URL and OAuth instead of a local process:

{
  "mcpServers": {
    "supabase": { "url": "https://mcp.supabase.com/mcp" },
    "sentry": { "url": "https://mcp.sentry.dev/mcp" },
    "vercel": { "url": "https://mcp.vercel.com" }
  }
}

On first use, the client will open a browser for OAuth login.

Top Server Categories

Code and repos

• GitHub MCP Server — Repos, PRs, issues, code search, file access.
• Filesystem MCP Server — Local file read/write within a sandbox (for clients without built-in file access).

Databases

• PostgreSQL MCP Server — Schema inspection and read-only SQL.
• MongoDB MCP Server — Explore collections, run queries, generate code.
• Redis MCP Server — Get, set, delete, and list keys.
• Supabase MCP Server — SQL, migrations, Edge Functions, docs (hosted, OAuth).

Product and docs

• Notion MCP Server — Search pages, read databases, create and update content.
• Linear MCP Server — Search and update issues.
• Jira MCP Server — Search issues, create and update tickets.

Design and deployment

• Figma MCP Server — Design context, variables, Code Connect, design-to-code.
• Vercel MCP Server — Projects, deployments, build logs, docs.
• Firebase MCP Server — Firestore, Auth, security rules.

Infrastructure and observability

• Sentry MCP Server — Issues, errors, Seer analysis.
• Kubernetes MCP Server — Pods, logs, Helm, generic resources.
• Docker MCP Server — Run MCP servers in isolated containers via the Docker gateway.

See the full MCP directory for all covered servers.

Tool Compatibility

Configuration format is similar across tools: mcpServers object with server name, command/args for stdio or url for remote. Check each tool's docs for exact paths and options.

Common Pitfalls

1. Config not loaded — Ensure the config file path is correct and restart the AI tool.
2. Missing env vars — Tokens in env must be set; avoid hardcoding secrets in committed files.
3. Wrong transport — Use command/args for stdio servers, url for HTTP. Mixing them causes failures.
4. Production access — Do not connect MCP servers to production databases or critical systems; use dev/staging.
5. Too many servers — Each server adds tools to the context. Enable only what you need for the current task.
6. Approval disabled — If your client supports it, keep manual approval for tool calls to avoid unintended mutations.

Security Recommendations

• Store API keys and tokens in environment variables, not in config files.
• Prefer read-only modes and scoped access (e.g., project-specific Supabase, repo-specific GitHub tokens).
• Use OAuth-hosted servers when available instead of long-lived tokens.
• Review tool call parameters before approving; agents can be prompted to run harmful commands.
• Isolate sensitive data; do not give agents access to PII or production databases.

Next Steps

• Configure MCP servers — Cross-tool setup guide.
• MCP ecosystem cookbook — Patterns and workflows.
• MCP directory — Browse all covered servers with setup instructions.

Related in This Cluster

• The MCP Revolution: How One Protocol Connects Every AI Tool
• MCP directory — All covered servers
• GitHub MCP Server
• Figma MCP Server
• Supabase MCP Server

Related guides: AI coding agents explained | How to use Cursor | Directory