alonf/mcppythondemo/python-mcp-sampling-bridge

Python MCP Sampling Bridge

Use When

• A Python MCP server needs sampling/createMessage parity with another implementation
• The client should own the LLM runtime while the server orchestrates sampling

Pattern

1. On the server, call ctx.session.create_message(...).
2. On the client, pass an async sampling_callback into ClientSession(...); this automatically advertises sampling capability.
3. Keep the callback thin: translate MCP messages to your model input, call the model, return CreateMessageResult.
4. Treat sampled output as untrusted input. Validate it on the server, retry with validation errors when needed, and only then execute deterministic work.
5. For Linux diagnostics teaching flows, a constrained PATH or PATH | grep FIELD format is easier to validate than free-form shell commands.

Repo Example

• src/mcp_linux_diag_server/tools/linux_diagnostics.py
• src/mcp_linux_diag_server/client.py
• tests/test_linux_diagnostics.py