MCP Streamable HTTP Servers & Client

This project demonstrates how to build and interact with Model Context Protocol (MCP) streamable HTTP servers and clients in Python. It includes stateless servers, a Google OAuth–protected server, and a Gemini-powered ADK client capable of interacting with MCP toolsets.

1️⃣ Stateless Streamable Servers

Location: streamable_http_server/1-stateless-streamable/

These are stateless, streamable HTTP servers built using the Model Context Protocol (MCP). Stateless means no memory or session is retained across tool calls.

server1.py: Provides add_numbers and subtract_numbers tools.
server2.py: Provides multiply_numbers and divide_numbers tools.
main.py: Launchpad script to run either server1 or server2 from CLI.

🚀 Getting Started

Create a virtual environment from the root directory

# macOS / Linux
uv venv
source ./.venv/bin/activate

# Windows (PowerShell)
uv venv
.venv\scripts\activate

Install requirements with uv
```
uv sync --all-groups
```

Run a Server

Run server1 (Add + Subtract):

uv run --active streamable_http_server/1-stateless-streamable/main.py --server server1

Run server2 (Multiply + Divide):

uv run --active streamable_http_server/1-stateless-streamable/main.py --server server2

2️⃣ Google OAuth–Protected Server

Location: streamable_http_server/2-google-oauth-simple-server/

This server demonstrates the OAuth Proxy pattern with Google as the upstream provider. It protects an MCP server behind Google OAuth 2.0, allowing MCP clients to authenticate dynamically using DCR (Dynamic Client Registration), PKCE, and loopback redirect URIs.

server.py: MCP Resource Server acting as an OAuth Proxy to Google.
README.md: Detailed explanation of setup, environment variables, and flow.

3️⃣ MCP Streamable HTTP Client

Location: streamable_http_client/

This is an educational project that demonstrates how to connect to a Model Context Protocol (MCP) streamable HTTP server, discover tools from the server, and interact with those tools using a Google ADK agent powered by Google Gemini.

⚙️ Setup Instructions

Create a virtual environment

# macOS / Linux
uv venv
source ./.venv/bin/activate

# Windows (PowerShell)
uv venv
.venv\scripts\activate

Install dependencies
```
uv sync --all-groups
```
Set environment variables

Create a .env file inside streamable_http_client:
```
GOOGLE_API_KEY=your-google-api-key
```

Configure MCP Servers

Edit the streamable_http_client/theailanguage_config.json file:

{
  "mcpServers": {
    "server1": {
      "type": "http",
      "url": "http://localhost:3000/mcp"
    },
    "server2": {
      "type": "http",
      "url": "http://localhost:3001/mcp"
    },
    "terminal": {
      "type": "stdio",
      "command": "/Users/theailanguage/.local/bin/uv",
      "args": [
          "--directory", "/Users/theailanguage/mcp/mcp_stremable_http/stdio_server/1-terminal-server",
          "run",
          "terminal_server.py"
      ]
    }
  }
}

Run the Client
```
uv run universal_client/1-google-adk-gemini-mcp-client/cmd.py
```
This launches an interactive command-line chat loop, connects to MCP servers via HTTP or STDIO, and interacts with the Gemini-powered ADK agent using tools discovered from each server.

4️⃣ Google OAuth–Protected Client

Location: universal_client/3-google-oauth-simple-client/

This client demonstrates how to authenticate against the Google OAuth–protected MCP server using the OAuth Proxy pattern. It:

Handles loopback redirect URIs.
Supports DCR + PKCE automatically.
Interacts with the protected tools (get_time, get_user_info).

Run with:

source .venv/bin/activate
uv run ./universal_client/3-google-oauth-simple-client/client.py

5️⃣ Coming Soon

🧠 Stateful Streamable Server

A stateful, streamable HTTP server using MCP that maintains state across tool invocations and enables resumable event streams.

🔧 Claude Desktop Integration

If you want to integrate these MCP servers with Claude Desktop, use the following config:

{
  "mcpServers": {
    "server1": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:3000/mcp"
      ]
    },
    "server2": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "http://localhost:3001/mcp"
      ]
    }
  }
}

Save this as claude_desktop_config.json. Warning! - This uses a third party package called mcp-remote that is not an official Anthropic or Claude package

📜 License

This repository and the code within are licensed under the GNU General Public License v3.0. See the LICENSE file for full details.

Built with ❤️ by The AI Language to teach and demonstrate how to create streamable MCP servers and agents in Python using FastMCP, Pydantic, and ADK.

流式HTTP