daBongo LMS AI Training Courses

Lesson Content

What sampling is.

In standard MCP flow, Claude (the model) calls tools on the server. Sampling reverses this: the server requests a completion from the model via the host client. The server does not have direct API access – it asks the host to make an LLM request on its behalf and return the result.

This enables servers to use AI capabilities without managing API keys, billing, or model selection. The host's existing model connection is reused.

When to use sampling.

Use sampling in a server when:

• A tool handler needs to process data with AI assistance before returning a result
• The server needs to make a decision that benefits from language model reasoning
• You want server-side AI processing without adding API access to the server itself

Do not use sampling when a direct API call from the server is simpler and clearer – sampling adds a round-trip through the host and is appropriate for cases where sharing the host's model connection is the right architecture.

Implementing sampling (Python pattern).

@app.call_tool()
async def call_tool(name: str, arguments: dict):
if name == "analyze_and_summarize":
raw_data = fetch_data(arguments["source"])

# Request a completion from the host via sampling
sampling_result = await app.request_sampling(
messages=[
{"role": "user", "content": f"Summarize this data: {raw_data}"}
],
max_tokens=200
)

return [types.TextContent(type="text", text=sampling_result.content[0].text)]

Verify current SDK method signatures for sampling at modelcontextprotocol.io/sdk – the pattern may vary between SDK versions.

Human-in-the-loop via sampling.

Sampling passes through the host – hosts can (and often do) surface sampling requests to the user for approval before sending to the model. This creates a human-in-the-loop mechanism for AI-driven server decisions. Design sampling requests with this in mind: if your server makes sensitive decisions via sampling, the human-approval path is a feature, not an obstacle.