python-mcp-tool-bridge

Python MCP Tool Bridge Skill

Use When

You need a Python lecture/demo client that lets an LLM call MCP tools without waiting for server-side sampling support.

Pattern

1. Start the MCP server locally with StdioServerParameters(command=sys.executable, args=["-m", server_module]).
2. Connect with stdio_client(...) + ClientSession, then initialize() and list_tools().
3. Convert each MCP Tool into an OpenAI function tool by reusing name, description, and inputSchema.
4. Run a chat-completions loop:
   • send system + user messages plus tool schema
   • if the model returns tool calls, execute them through session.call_tool(...)
   • feed serialized tool results back as role="tool"
   • stop when the model returns plain assistant text
5. Keep config explicit via env vars for endpoint, API key, deployment, and API version.

Runtime Shape Check

• Before adding Azure AI Foundry project/runtime abstractions, inspect the live reference client code that actually runs the lecture flow.
• If that runnable path uses AzureOpenAIClient(endpoint, credential).GetChatClient(deploymentName) or the equivalent endpoint + deployment + credential shape, keep the Python bridge on the simpler Azure OpenAI runtime path.
• Only switch the Python client to a Foundry project client when the reference app itself is using a project endpoint/client in its active runnable path.

Testing Pattern

• For live validation in WSL, first verify az account show succeeds, then run python3 -m mcp_linux_diag_server.client --json --prompt "..."; the [tool] ... trace plus final JSON answer is the fastest proof that Azure OpenAI and the MCP bridge both worked.
• Unit-test config parsing independently.
• Unit-test MCP-tool → OpenAI-tool translation as a pure function.
• Unit-test the agent loop with a fake model and fake MCP session so no external credentials are required.
• Keep smoke tests focused on server startup plus client configuration guardrails.

Reference Files