Sampling

Das Model Context Protocol (MCP) bietet einen standardisierten Weg für Server, LLM- Sampling (“Vervollständigungen” oder “Generierungen”) von Sprachmodellen über Clients anzufordern. Dieser Ablauf ermöglicht es Clients, die Kontrolle über Modellzugriff, -auswahl und -berechtigungen zu behalten, während Server KI-Fähigkeiten nutzen können—ohne dass Server-API-Schlüssel erforderlich sind. Server können text- oder bildbasierte Interaktionen anfordern und optional Kontext von MCP-Servern in ihre Prompts einbeziehen.

Benutzerinteraktionsmodell

Sampling in MCP ermöglicht es Servern, agentische Verhaltensweisen zu implementieren, indem LLM-Aufrufe verschachtelt innerhalb anderer MCP-Server-Funktionen auftreten können.

Implementierungen sind frei, Sampling über jedes Schnittstellenmuster bereitzustellen, das ihren Bedürfnissen entspricht—das Protokoll selbst schreibt kein spezifisches Benutzerinteraktionsmodell vor.

Fähigkeiten

Clients, die Sampling unterstützen, MÜSSEN die sampling-Fähigkeit während der Initialisierung deklarieren:

{
  "capabilities": {
    "sampling": {}
  }
}

Protocol Messages

Creating Messages

To request a language model generation, servers send a sampling/createMessage request:

Request:

{
  "jsonrpc": "2.0",
  "id": 1,
  "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
  }
}

Response:

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "role": "assistant",
    "content": {
      "type": "text",
      "text": "The capital of France is Paris."
    },
    "model": "claude-3-sonnet-20240307",
    "stopReason": "endTurn"
  }
}

Message Flow

  sequenceDiagram
    participant Server
    participant Client
    participant User
    participant LLM

    Note over Server,Client: Server initiates sampling
    Server->>Client: sampling/createMessage

    Note over Client,User: Human-in-the-loop review
    Client->>User: Present request for approval
    User-->>Client: Review and approve/modify

    Note over Client,LLM: Model interaction
    Client->>LLM: Forward approved request
    LLM-->>Client: Return generation

    Note over Client,User: Response review
    Client->>User: Present response for approval
    User-->>Client: Review and approve/modify

    Note over Server,Client: Complete request
    Client-->>Server: Return approved response

Data Types

Messages

Sampling messages can contain:

Text Content

{
  "type": "text",
  "text": "The message content"
}

Image Content

{
  "type": "image",
  "data": "base64-encoded-image-data",
  "mimeType": "image/jpeg"
}

Model Preferences

Model selection in MCP requires careful abstraction since servers and clients may use different AI providers with distinct model offerings. A server cannot simply request a specific model by name since the client may not have access to that exact model or may prefer to use a different provider’s equivalent model.

To solve this, MCP implements a preference system that combines abstract capability priorities with optional model hints:

Capability Priorities

Servers express their needs through three normalized priority values (0-1):

• costPriority: How important is minimizing costs? Higher values prefer cheaper models.
• speedPriority: How important is low latency? Higher values prefer faster models.
• intelligencePriority: How important are advanced capabilities? Higher values prefer more capable models.

Model Hints

While priorities help select models based on characteristics, hints allow servers to suggest specific models or model families:

• Hints are treated as substrings that can match model names flexibly
• Multiple hints are evaluated in order of preference
• Clients MAY map hints to equivalent models from different providers
• Hints are advisory—clients make final model selection

For example:

{
  "hints": [
    { "name": "claude-3-sonnet" }, // Prefer Sonnet-class models
    { "name": "claude" } // Fall back to any Claude model
  ],
  "costPriority": 0.3, // Cost is less important
  "speedPriority": 0.8, // Speed is very important
  "intelligencePriority": 0.5 // Moderate capability needs
}

The client processes these preferences to select an appropriate model from its available options. For instance, if the client doesn’t have access to Claude models but has Gemini, it might map the sonnet hint to gemini-1.5-pro based on similar capabilities.

Error Handling

Clients SHOULD return errors for common failure cases:

Example error:

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": {
    "code": -1,
    "message": "User rejected sampling request"
  }
}

Security Considerations

1. Clients SHOULD implement user approval controls
2. Both parties SHOULD validate message content
3. Clients SHOULD respect model preference hints
4. Clients SHOULD implement rate limiting
5. Both parties MUST handle sensitive data appropriately