[Question] Discrepancy between MCP and Python Lib, and missing complete OpenAI SDK example?

[Slipstream-Max]

​Hi team,
​I am currently exploring the code-mode project and comparing code-mode-mcp with the Python library. I noticed some structural differences and have a few questions regarding the implementation:
​1. Discrepancy in tool count
The MCP server exposes 7 tools (including register_manual, get_required_keys_for_tool, etc.), while the Python library only provides 3 core capabilities. Please clarify the design philosophy behind this difference.
​2. Missing list_tools and tool_info in Python Lib
In standard Agent architectures, the LLM usually relies on tools like list_tools and tool_info to dynamically discover and understand available APIs. Since these specific functions are missing in the Python library, I would like to know the officially recommended approach for handling dynamic tool discovery and parameter inspection in the Python SDK.
​3. Request for a complete OpenAI SDK integration example
Currently, the documentation mostly demonstrates how to construct the system prompt, but it lacks a complete demonstration of the actual tool-calling loop.
To help developers integrate this smoothly, it would be best if you could provide a ready-to-use OpenAI tools schema (the JSON schema definition) for the available Python functions. Additionally, a minimal complete example showing how to wire these tools into a standard native OpenAI SDK while loop would be incredibly helpful.
​Thanks for your time and for the great project!

[Docat0209]

Hi! Great questions. The discrepancies you're seeing between the MCP server and Python library design are super common—glad you asked.

Here's some context that might help:

1. Tool count difference — The MCP server likely exposes meta-tools like register_manual and get_required_keys for agent-level introspection. The Python library keeps it minimal (3 core capabilities) because it's designed as a low-level SDK. Different design philosophies for different use cases.

2. Dynamic tool discovery — For standard LLM tool calling, you're usually right that you need list_tools and tool_info. However, with MCP, the agent discovers tools at initialization time via the protocol, not dynamically per-call. Check the MCP spec's tools/list resource type—it's the standard way agents discover tools.

3. OpenAI SDK integration example — This is a great doc request! A worked example with OpenAI's function_calling + MCP would be super valuable. You might also check if the project has any integration examples in their tests or examples/ folder.

I'd suggest opening a documentation PR or reaching out to the maintainers—they'd probably love a contribution here. The code-mode project seems to be building something important, and good examples make all the difference.