Deploy a Self-Hosted MCP Server on a VPS

Third-party hosted MCP services introduce an external dependency: your data, prompts and contexts pass through infrastructure you do not control. On a VPS, the MCP server runs in your network, with your access rules and your chain of trust. This is the same logic as self-hosting a Git repository or a Docker registry: autonomy takes precedence over convenience.

You need a VPS with root access, a recent Linux distribution (Ubuntu 22.04 LTS or Debian 12 are recommended), Node.js 18+ or Python 3.11+ depending on the implementation chosen, and a domain name pointing to your server's IP. A plan with at least 1 vCPU and 2 GB of RAM is suitable for a moderately used MCP server. For heavy loads or multiple MCP servers in parallel, plan for 2 vCPU and 4 GB.

An MCP server exposed on the Internet is an attack surface. Four measures are non-negotiable before connecting an AI client to this server in production.

Bearer token authentication. Enter a unique token per client in the Authorization header. Immediately revoke a compromised token and generate a new one with openssl rand -hex 32. Do not reuse the same token for multiple clients: you will lose traceability.

nginx rate-limiting. Add to your vhost: limit_req_zone $binary_remote_addr zone=mcp:10m rate=30r/m; limit_req zone=mcp burst=10 nodelay;. This limits to 30 requests per minute per IP, with a burst of 10. Adjust according to your actual usage.

Mandatory HTTPS and HSTS. After installing Certbot, add in the HTTPS server block: add_header Strict-Transport-Security "max-age=31536000" always;. This prevents clients from connecting over plain HTTP even if they try.

Token rotation. Schedule a rotation every 90 days via cron or your secrets management tool. Document each active token and its associated client in a private registry (not in the code repository).

Once the server is online and secured, connecting from an AI client takes less than two minutes.

Claude Desktop. Open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Add an entry in the mcpServers key:

{
  "mcpServers": {
    "my-vps": {
      "url": "https://your-domain.com",
      "headers": {
        "Authorization": "Bearer YOUR_SECRET_TOKEN"
      }
    }
  }
}

Restart Claude Desktop. The server appears in the list of available tools.

Cursor. In settings (Cursor > Settings > MCP), click 'Add server', choose the HTTP type, enter the URL and add the Authorization header. Cursor automatically injects MCP tools into the context of each conversation.

Verification. Ask the agent to list available tools. If the filesystem server is active, the response should include read_file, write_file and list_directory. The absence of these tools indicates a configuration error — check the nginx logs first.

1. MCP transport error: Failed to fetch
This message appears in Claude Desktop or Cursor when the client cannot reach the configured URL. Common causes: the systemd service is not running (systemctl status mcp-server), nginx has not been reloaded after modifying the vhost (nginx -t && systemctl reload nginx), or a firewall is blocking port 443 (ufw status).

2. Connection refused on 127.0.0.1:3000
The MCP server did not start or crashed at startup. Check the systemd logs: journalctl -u mcp-server -n 50. Common causes are an invalid JSON configuration file, a missing npm dependency (npm install to re-run in the project directory) or a port already in use (ss -tlnp | grep 3000).

3. Request timed out after 30000ms
The server responds but too slowly. On a filesystem tool, this happens if the directory being scanned is too large or the disk is full (df -h). For a postgres tool, check that the query is not doing a full scan on a table without an index. Increase the nginx timeout as a last resort: proxy_read_timeout 120s;.

4. 401 Unauthorized systematically
The token sent by the client does not match the one expected by nginx. Check the token accuracy in the client configuration (trailing space, quotes accidentally included). If you have just renewed the token server-side, restart nginx and reload the AI client configuration.

An MCP server is one component among many in a self-hosted AI environment. On the same VPS, you can run a local model via Ollama, a vector database like Qdrant or Chroma, and an orchestrator like n8n or Flowise — each connected to the MCP server via dedicated tools. ServOrbit offers ready-to-use VPS templates for these environments, installable in a few clicks from the Marketplace, on infrastructure hosted in European data centres with dedicated IPv4 and root access.