Deploy an MCP Server on AWS EC2

A local MCP server only helps you. To share it — or to give Claude Desktop a tool that runs in the cloud — you deploy it to a server. This lesson hosts the research assistant server on an Amazon Linux EC2 instance, serving it over the Streamable HTTP transport so any MCP client can reach it.

Two changes adapt the server for the cloud: it embeds with OpenAI (no local GPU needed) and it binds to 0.0.0.0 so it listens on the instance's public address.

Launch the EC2 instance

In the AWS console, launch an Amazon Linux EC2 instance and:

1. Create or choose a key pair (download the .pem file, e.g. mcp.pem).
2. In the security group, allow inbound SSH (port 22) from your IP and a custom TCP rule for port 8000 (the MCP server port).

Connect over SSH

Use the .pem key to connect. Replace the host with your instance's public DNS from the EC2 console.

From Windows (PowerShell / Command Prompt):

# If the .pem file is in the current directory
ssh -i "mcp.pem" ec2-user@your-ec2-public-dns.compute.amazonaws.com

On Linux/macOS: the same command works, but you may first need to tighten the key's permissions:

chmod 400 mcp.pem
ssh -i "mcp.pem" ec2-user@your-ec2-public-dns.compute.amazonaws.com

Install Anaconda and uv on the server

Once connected, update the system and install Anaconda.

# 1. Update the system
sudo yum update -y

# 2. Download the Anaconda installer
curl -O https://repo.anaconda.com/archive/Anaconda3-2025.06-0-Linux-x86_64.sh

# 3. Run the installer
bash Anaconda3-2025.06-0-Linux-x86_64.sh

Follow the prompts, accept the license, and use the default install location /home/ec2-user/anaconda3. When asked about shell initialization, choose "no" initially, then add it to PATH manually:

export PATH="/home/ec2-user/anaconda3/bin:$PATH"
echo 'export PATH="/home/ec2-user/anaconda3/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

Initialize conda and install uv:

conda init bash
source ~/.bashrc
pip install uv

Verify the toolchain:

conda --version
python --version
uv --version

Adapt the server for the cloud

Copy your server folder to the instance (with scp, git clone, or by pasting the file) and adjust two things in server.py.

First, embed with OpenAI instead of Ollama — a cloud server has no local model:

from langchain_openai import OpenAIEmbeddings
from dotenv import load_dotenv

load_dotenv()

EMBED_MODEL = "text-embedding-3-small"
embeddings = OpenAIEmbeddings(model=EMBED_MODEL)

Second, serve over HTTP bound to all interfaces so the instance's public IP can receive requests:

if __name__ == "__main__":
    mcp.run(transport="http", host="0.0.0.0")

Create a .env on the server with your key:

OPENAI_API_KEY=your-openai-api-key

Install dependencies and run it:

uv sync
uv run python server.py

The server is now reachable at http://your-ec2-public-dns.compute.amazonaws.com:8000/mcp.

Connect Claude Desktop to the remote server

On your laptop, point Claude Desktop at the remote server using the mcp-remote bridge in claude_desktop_config.json:

{
  "mcpServers": {
    "remote-research-server": {
      "command": "npx",
      "args": ["mcp-remote", "http://your-ec2-public-dns.compute.amazonaws.com:8000/mcp", "--allow-http"]
    }
  }
}

Restart Claude Desktop, and your cloud-hosted research tools appear alongside your local ones.

Recap

• Launch Amazon Linux EC2, open ports 22 and 8000, and SSH in as ec2-user.
• Install Anaconda + uv, then uv sync your server's dependencies.
• For the cloud, embed with OpenAI and run mcp.run(transport="http", host="0.0.0.0").
• Keep the process alive with tmux/systemd, and reach it from Claude Desktop via mcp-remote.

A server tied to one EC2 box is hard to share widely. The final step is packaging your server so anyone can install it with one command — see Publish Your MCP Server on PyPI.