▶ REVERSE AI PATTERN
Server sends prompt to model mid-tool execution for AI-driven workflows
async def sampleLLM(prompt: str) -> str:
"""Server requests completion from the connected model"""
response = await server.request_completion(
messages=[{"role": "user", "content": prompt}],
max_tokens=1000
)
return response.content
▶ USE CASE PATTERNS
🤖 SUMMARIZATION
Generate summaries during data processing
🔧 CODE GENERATION
AI-assisted tool operations
👤 USER EMULATION
Simulate user responses for testing
▶ SAMPLING WORKFLOW ARCHITECTURE
┌──────────────────────────────────────────────────────────────┐
│ SAMPLING WORKFLOW │
├──────────────────────────────────────────────────────────────┤
│ │
│ 1. Tool Execution 2. Server Decision 3. Model Call │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ User calls │────►│ Server needs│─────►│ Model │ │
│ │ sampleLLM │ │ AI assist │ │ generates │ │
│ └─────────────┘ └─────────────┘ │ response │ │
│ ▲ └─────────────┘ │
│ │ │ │
│ │ 4. Validation 5. Return│ │
│ ┌─────────────┐ ┌─────────────┐ ▼ │
│ │ Final │◄────┤ Server │◄────┌─────────────┐ │
│ │ Response │ │ validates │ │ AI Response │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │
│ Critical: Always validate AI output before trusting actions│
└──────────────────────────────────────────────────────────────┘
🎮 LIVE DEMO: SAMPLELLM TOOL
DEMONSTRATION:
sampleLLM tool in Everything Server triggers completions
$ mcp-inspector
> Call sampleLLM tool
> Prompt: "Summarize the key benefits of MCP in 3 bullet points"
[TOOL] sampleLLM called with prompt
[SERVER] Requesting completion from connected model...
[MODEL] Processing prompt and generating response...
Response:
┌─────────────────────────────────────────────┐
│ • Modular Integration: MCP enables M+N │
│ connections instead of M×N complexity │
│ │
│ • Standardized Protocol: Unified interface │
│ for tools, resources, and prompts │
│ │
│ • AI-Native Design: Built specifically for │
│ LLM interaction patterns and workflows │
└─────────────────────────────────────────────┘
[VALIDATION] Server validates response format and content
[RETURN] Validated AI-generated summary returned to user
Key Observations:
- Bidirectional AI - server can request model assistance
- Validation layer - never trust AI output blindly
- Workflow integration - AI becomes part of tool logic
- Context preservation - model has access to conversation context
▶ ADVANCED APPLICATIONS
├── Data Analysis │ AI-powered data insights and patterns
├── Content Generation │ Dynamic content creation workflows
├── Decision Support │ AI-assisted decision making
├── Quality Assurance │ Automated testing and validation
└── User Experience │ Personalized responses and recommendations
▶ SAFETY CONSIDERATIONS
✅ ALWAYS VALIDATE
Never trust AI output for critical operations
🔒 SANDBOX EXECUTION
Isolate AI-generated code or commands
📊 AUDIT TRAILS
Log all AI interactions for debugging