MCP Server & Client w/ Azure OpenAI & OpenAI

A minimal server/client application implementation utilizing the Model Context Protocol (MCP) and Azure OpenAI.
1. The MCP server is built with FastMCP.
2. Playwright is an an open source, end to end testing framework by Microsoft for testing your modern web applications.
3. The MCP response about tools will be converted to the OpenAI function calling format.
4. The bridge that converts the MCP server response to the OpenAI function calling format customises the MCP-LLM Bridge implementation.
5. To ensure a stable connection, the server object is passed directly into the bridge.
6. The client_bridge supports both in-process and external (stdio) MCP server connections, enabling reuse by different clients (e.g., Claude Code, VS Code, custom scripts).

Model Context Protocol (MCP)

Model Context Protocol (MCP) MCP (Model Context Protocol) is an open protocol that enables secure, controlled interactions between AI applications and local or remote resources.

Community Resources

FastMCP: The fast, Pythonic way to build MCP servers.
Chat MCP: MCP client
MCP-LLM Bridge: MCP implementation that enables communication between MCP servers and OpenAI-compatible LLMs

During the development phase in December 2024, the Python project should be initiated with 'uv'. Other dependency management libraries, such as 'pip' and 'poetry', are not yet fully supported by the MCP CLI.

Rename .env.template to .env, then fill in the values in .env for Azure OpenAI:

AZURE_OPEN_AI_ENDPOINT=
AZURE_OPEN_AI_API_KEY=
AZURE_OPEN_AI_DEPLOYMENT_MODEL=
AZURE_OPEN_AI_API_VERSION=

Install uv for python library management
```
pip install uv
uv sync
```
Execute python chatgui.py
- The sample screen shows the client launching a browser to navigate to the URL.

Using with External Clients

The MCP server can be used by external clients (Claude Desktop, VS Code, Claude Code, etc.) via mcp.json configuration.

Claude Desktop / Claude Code

Add to your claude_desktop_config.json (Claude Desktop) or .claude/mcp.json (Claude Code):

{
  "mcpServers": {
    "browser-navigator": {
      "command": "uv",
      "args": ["run", "fastmcp", "run", "./server/browser_navigator_server.py:app"],
      "cwd": "/path/to/mcp-aoai-web-browsing",
      "env": {
        "AZURE_OPEN_AI_ENDPOINT": "...",
        "AZURE_OPEN_AI_API_KEY": "...",
        "AZURE_OPEN_AI_DEPLOYMENT_MODEL": "...",
        "AZURE_OPEN_AI_API_VERSION": "..."
      }
    }
  }
}

VS Code

Add to .vscode/mcp.json in your workspace:

{
  "servers": {
    "browser-navigator": {
      "command": "uv",
      "args": ["run", "fastmcp", "run", "./server/browser_navigator_server.py:app"],
      "cwd": "${workspaceFolder}",
      "env": {
        "AZURE_OPEN_AI_ENDPOINT": "...",
        "AZURE_OPEN_AI_API_KEY": "...",
        "AZURE_OPEN_AI_DEPLOYMENT_MODEL": "...",
        "AZURE_OPEN_AI_API_VERSION": "..."
      }
    }
  }
}

Using the Bridge Programmatically (stdio)

The client_bridge also supports connecting to external MCP servers via stdio from Python:

from client_bridge import BridgeConfig, MCPServerConfig, BridgeManager
from client_bridge.llm_config import get_default_llm_config
 
config = BridgeConfig(
    server_config=MCPServerConfig(
        command="uv",
        args=["run", "fastmcp", "run", "./server/browser_navigator_server.py:app"],
    ),
    llm_config=get_default_llm_config(),
    system_prompt="You are a helpful assistant.",
)
 
async with BridgeManager(config) as bridge:
    response = await bridge.process_message("Navigate to https://example.com")

Using Standard OpenAI (non-Azure)

from client_bridge.llm_config import get_openai_llm_config
 
config = BridgeConfig(
    mcp=server,
    llm_config=get_openai_llm_config(),
)

Set environment variables:

OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-...

Direct Tool Execution

For clients that manage their own LLM loop, the bridge exposes tool metadata and direct execution:

async with BridgeManager(config) as bridge:
    tools = bridge.get_tools()  # OpenAI function calling format
    result = await bridge.execute_tool("playwright_navigate", {"url": "https://example.com"})

w.r.t. 'stdio'

stdio is a transport layer (raw data flow), while JSON-RPC is an application protocol (structured communication). They are distinct but often used interchangeably, e.g., "JSON-RPC over stdio" in protocols.

Tool description

@self.mcp.tool()
async def playwright_navigate(url: str, timeout=30000, wait_until="load"):
    """Navigate to a URL.""" -> This comment provides a description, which may be used in a mechanism similar to function calling in LLMs.
 
# Output
Tool(name='playwright_navigate', description='Navigate to a URL.', inputSchema={'properties': {'url': {'title': 'Url', 'type': 'string'}, 'timeout': {'default': 30000, 'title': 'timeout', 'type': 'string'}

Tip: uv

features

uv run: Run a script.
uv venv: Create a new virtual environment. By default, '.venv'.
uv add: Add a dependency to a script
uv remove: Remove a dependency from a script
uv sync: Sync (Install) the project's dependencies with the environment.

Tip

taskkill command for python.exe

taskkill /IM python.exe /F

Visual Code: Python Debugger: Debugging with launch.json will start the debugger using the configuration from .vscode/launch.json.

kimtth/mcp-aoai-web-browsing