ما الذي يتيحه Sampling؟

Usually the model calls your tools. Sampling flips that: your server asks the model to generate text for it. 🔄

With sampling, your server can use the host’s LLM without ever holding an API key. The client owns the model access.

Because the client runs the completion, your server needs no provider credentials. That keeps secrets out of your code entirely.

Sampling lets an LLM call happen nested inside a tool. Your tool can think with the model mid-task, not just return data.

Imagine a tool that fetches a report, then asks the model to summarize it. The summarizing happens through sampling.

The client picks the model and pays for it, so the host decides cost, limits, and which model runs your request.

Think of sampling as the server saying: I have data, you have a brain. It delegates the reasoning step back to the model.

A sampling request can ask the model for text, and clients may also support image or audio content in messages.

Sampling only works if the client declares the sampling capability at startup. Not every host enables it.

Sampling unlocks smart, self-contained servers: data tools that can reason, classify, and write — powered by the user’s own model.

Because servers can trigger model calls, there must be a human able to approve them. You’ll see that safety flow soon.

Quick gut check on what sampling really does.

Sampling flips control: your server asks the host’s model to think for it — no keys, opt-in, and always with a human able to approve. ✅