MCP Sampling: Status and What Replaced It

Status in the 2026-07-28 spec: deprecated (still functional; eligible for removal in the first spec released on or after 28 July 2027)

Sampling is an MCP feature that lets a server ask the client's AI model to generate text for it. Instead of holding its own model API key, the server sends a prompt to the AI app, and the app runs it through whatever model the user is already using.

The 2026-07-28 MCP specification deprecated Sampling. It still works, but new projects should not adopt it. The spec's suggested replacement is simple: "integrate directly with LLM provider APIs." This page explains what Sampling did, why it is being retired, and how to move off it.

What Sampling was for

Some servers need a bit of AI inside their own logic. A server might want to summarize a document, classify a ticket, or decide the next step in a workflow. Sampling let the server borrow the client's model for that.

The Sampling page of the spec says it lets clients "maintain control over model access, selection, and permissions." Servers get AI capabilities "with no server API keys necessary." The spec also says there should always be "a human in the loop with the ability to deny sampling requests."

Sampling has been part of MCP since the first published version, 2024-11-05.

How it works in 2026-07-28

Sampling still functions during the deprecation window. Like elicitation, it now travels through Multi Round-Trip Requests. The server cannot send its own request to the client, so it returns an input_required result containing a sampling/createMessage request. The client gets the user's approval, runs the model, and retries the original call with the model's answer.

Here is the request, adapted from the spec example (we removed the deprecated includeContext field):

{
  "method": "sampling/createMessage",
  "params": {
    "messages": [
      {
        "role": "user",
        "content": { "type": "text", "text": "What is the capital of France?" }
      }
    ],
    "modelPreferences": {
      "hints": [{ "name": "claude-3-sonnet" }],
      "intelligencePriority": 0.8,
      "speedPriority": 0.5
    },
    "systemPrompt": "You are a helpful assistant.",
    "maxTokens": 100
  }
}

A few details are worth knowing:

• Model choice stays with the client. The server can give hints and priorities for cost, speed and intelligence. The spec says "hints are advisory" and the client makes the final model choice.
• maxTokens is required. The client must respect it but may ignore temperature and similar settings.
• Tool use is possible. Since 2025-11-25, a server can pass tools and toolChoice so the model can call tools during sampling. The client must declare sampling.tools first.

Why it was deprecated

The changelog lists the deprecation of Roots, Sampling and Logging under SEP-2577. The spec pages state the outcome and the migration path. They do not give a long rationale, so we will not invent one.

What we can say from the spec itself: Sampling depends on the server reaching back to the client mid-request. The 2026-07-28 release rebuilt MCP so that servers never start requests. Sampling survives only by being squeezed into the retry pattern. Calling a model provider directly is a plainer design that the protocol does not need to carry.

The deprecation timeline

The deprecated features registry lists Sampling as:

The registry notes that earliest removal is when a feature becomes "eligible" for removal. The actual removal is a Core Maintainer decision and "may happen later." The general rules are in the feature lifecycle policy and our guide to MCP versioning and deprecation.

The includeContext values "thisServer" and "allServers" were soft-deprecated in 2025-11-25. They are now formally deprecated and will be removed no later than Sampling. Omit the field or use "none".

What changed across spec versions

How to migrate

1. Find where you use it. Search your server code for createMessage or your SDK's sampling helper.
2. Pick a model provider. Call it directly from your server with your own API key, stored as a secret.
3. Budget for cost and rate limits. With Sampling, the user's AI app paid for the model call. Now your server does.
4. Keep humans in the loop where it matters. If a sampled answer led to an action, consider asking the user with elicitation before acting.
5. Stop advertising reliance on it. Remove sampling-only code paths once your server no longer needs clients to support it.

Common mistakes

• Starting a new server on Sampling. The spec says new implementations "SHOULD NOT adopt it."
• Copying spec examples that use includeContext: "thisServer". That value is deprecated.
• Assuming you get a specific model. Hints are advice. The client can map them to any model.
• Forgetting the cost shift. Direct model calls cost your business money and need limits against runaway usage.

Next step

Sampling was one of three features deprecated together. See what happened to the others in MCP Roots: status and what replaced them, or return to the MCP specification reference. If you need help moving a server off Sampling, our MCP development team can scope it.

Frequently asked questions