The following article originally appeared on Block’s blog and is being republished here with the author’s permission.
If you’ve been following MCP, you’ve probably heard about tools which are functions that let AI assistants do things like read files, query databases, or call APIs. But there’s another MCP feature that’s less talked about and arguably more interesting: sampling.
Sampling flips the script. Instead of the AI calling your tool, your tool calls the AI.
Let’s say you’re building an MCP server that needs to do something intelligent like summarize a document, translate text, or generate creative content. You have three options:
Option 1: Hardcode the logic. Write traditional code to handle it. This works for deterministic tasks, but falls apart when you need flexibility or creativity.
Option 2: Bake in your own LLM. Your MCP server makes its own calls to OpenAI, Anthropic, or whatever. This works, but now you’ve got API keys to manage and costs to track, and you’ve locked users into your model choice.
Option 3: Use sampling. Ask the AI that’s already connected to do the thinking for you. No extra API keys. No model lock-in. The user’s existing AI setup handles it.
How Sampling Works
When an MCP client like goose connects to an MCP server, it establishes a two-way channel. The server can expose tools for the AI to call, but it can also request that the AI generate text on its behalf.
Here’s what that looks like in code (using Python with FastMCP):
The ctx.sample() call sends a prompt back to the connected AI and waits for a response. From the user’s perspective, they just called a “summarize” tool. But under the hood, that tool delegated the hard part to the AI itself.
A Real Example: Council of Mine
Council of Mine is an MCP server that takes sampling to an extreme. It simulates a council of nine AI personas who debate topics and vote on each other’s opinions.
But there’s no LLM running inside the server. Every opinion, every vote, every bit of reasoning comes from sampling requests back to the user’s connected LLM.
The council has nine members, each with a distinct personality:
- 🔧 The Pragmatist – “Will this actually work?”
- 🌟 The Visionary – “What could this become?”
- 🔗 The Systems Thinker – “How does this affect the broader system?”
- 😊 The Optimist – “What’s the upside?”
- 😈 The Devil’s Advocate – “What if we’re completely wrong?”
- 🤝 The Mediator – “How can we integrate these perspectives?”
- 👥 The User Advocate – “How will real people interact with this?”
- 📜 The Traditionalist – “What has worked historically?”
- 📊 The Analyst – “What does the data show?”
Each personality is defined as a system prompt that gets prepended to sampling requests.
When you start a debate, the server makes nine sampling calls, one for each council member:
That temperature=0.8 setting encourages diverse, creative responses. Each council member “thinks” independently because each is a separate LLM call with a different personality prompt.
After opinions are collected, the server runs another round of sampling. Each member reviews everyone else’s opinions and votes for the one that resonates most with their values:
The server parses the structured response to extract votes and reasoning.
One more sampling call generates a balanced summary that incorporates all perspectives and acknowledges the winning viewpoint.
Total LLM calls per debate: 19
- 9 for opinions
- 9 for voting
- 1 for synthesis
All of those calls go through the user’s existing LLM connection. The MCP server itself has zero LLM dependencies.
Benefits of Sampling
Sampling enables a new category of MCP servers that orchestrate intelligent behavior without managing their own LLM infrastructure.
No API key management: The MCP server doesn’t need its own credentials. Users bring their own AI, and sampling uses whatever they’ve already configured.
Model flexibility: If a user switches from GPT to Claude to a local Llama model, the server automatically uses the new model.
Simpler architecture: MCP server developers can focus on building a tool, not an AI application. They can let the AI be the AI, while the server focuses on orchestration, data access, and domain logic.
When to Use Sampling
Sampling makes sense when a tool needs to:
- Generate creative content (summaries, translations, rewrites)
- Make judgment calls (sentiment analysis, categorization)
- Process unstructured data (extract info from messy text)
It’s less useful for:
- Deterministic operations (math, data transformation, API calls)
- Latency-critical paths (each sample adds round-trip time)
- High-volume processing (costs add up quickly)
The Mechanics
If you’re implementing sampling, here are the key parameters:
The response object contains the generated text, which you’ll need to parse. Council of Mine includes robust extraction logic because different LLM providers return slightly different response formats:
Security Considerations
When you’re passing user input into sampling prompts, you’re creating a potential prompt injection vector. Council of Mine handles this with clear delimiters and explicit instructions:
This isn’t bulletproof, but it raises the bar significantly.
Try It Yourself
If you want to see sampling in action, Council of Mine is a great playground. Ask goose to start a council debate on any topic and watch as nine distinct perspectives emerge, vote on each other, and synthesize into a conclusion all powered by sampling.