Examples and Use Cases

This page demonstrates real-world usage of the CMF MCP Server with AI assistants, showing how to explore metadata, analyze lineage, and integrate with various AI platforms.

Basic Queries

Discovering Pipelines

User Query:

What pipelines are available in CMF?

AI Assistant Response:

The AI assistant will use the cmf_show_pipelines tool to retrieve all pipelines:

[
  {
    "cmfClient": "http://server:8080",
    "data": [
      {"name": "Test-env", "id": "1"},
      {"name": "training-pipeline", "id": "2"},
      {"name": "inference-pipeline", "id": "3"}
    ]
  }
]

The assistant then provides a human-readable response:

I found 3 pipelines in your CMF Server: 1. Test-env 2. training-pipeline 3. inference-pipeline

Exploring Executions

User Query:

Show me executions in the Test-env pipeline

AI Assistant Action:

The assistant calls cmf_show_executions(pipeline="Test-env") and presents the results:

The Test-env pipeline has the following executions: - train_model (UUID: abc123...) - evaluate_model (UUID: def456...) - data_processing (UUID: ghi789...)

Example: Querying pipelines and executions with an AI assistant

Lineage Analysis

Execution Lineage

Understanding execution relationships is crucial for debugging and reproducibility.

User Query:

What is the execution lineage for UUID starting with abc1?

AI Assistant Workflow:

1. Validates the pipeline exists using cmf_show_pipelines
2. Retrieves execution lineage using cmf_execution_lineage(pipeline="Test-env", selected_uuid="abc1")
3. Presents the lineage tree showing upstream and downstream executions

Example Response:

The execution lineage for abc1 shows: - Parent Execution: data_preprocessing (ghi789) - Current Execution: train_model (abc123) - Child Executions: evaluate_model (def456), deploy_model (jkl012)

Example: Execution lineage visualization showing execution dependencies

Artifact Lineage

Trace data flow through your pipeline to understand artifact provenance.

User Query:

Show me the artifact lineage for the Test-env pipeline

AI Assistant Action:

Calls cmf_artifact_lineage(pipeline="Test-env") and visualizes the artifact flow:

The artifact lineage shows: - Dataset → data_preprocessing → Processed_Dataset - Processed_Dataset → train_model → Model - Model + Test_Dataset → evaluate_model → Metrics

Example: Artifact lineage graph showing data flow through pipeline stages

Model Cards

Retrieve comprehensive model metadata for governance and analysis.

Basic Model Card Query

User Query:

What is the Model ID from Test-env?

AI Assistant Workflow:

1. Uses cmf_show_artifacts(pipeline="Test-env", artifact_type="Model")
2. Returns the model ID for further queries

Response:

I found a Model artifact in Test-env with ID: 42

Detailed Model Card

User Query:

What is the detailed model card of the Model in Test-env pipeline?

AI Assistant Workflow:

1. First retrieves the model ID using cmf_show_artifacts
2. Then calls cmf_show_model_card(model_id="42")
3. Presents all four sections of the model card

Example Response Structure:

Model Card for Model ID 42

Model Data: - Name: trained_model.pkl - Version: 1.0 - Framework: TensorFlow - Size: 25.3 MB

Model Execution: - Training time: 2025-02-09 10:30:00 - Parameters: learning_rate=0.001, batch_size=32 - Training duration: 3.5 hours

Model Input Artifacts: - training_dataset.csv (Dataset) - validation_dataset.csv (Dataset)

Model Output Artifacts: - training_metrics.json (Metrics) - model_checkpoint.pkl (Model)

Example: Querying model card information

Example: Model card results showing comprehensive model metadata

Multi-Server Queries

Query across development, staging, and production environments simultaneously.

User Query:

Compare model versions between dev and production CMF servers

AI Assistant Workflow:

Queries both servers using cmf_show_artifacts with cmfClient_instances parameter:

cmf_show_artifacts(
    pipeline="training-pipeline",
    artifact_type="Model",
    cmfClient_instances=[
        "http://dev-server:8080",
        "http://prod-server:8080"
    ]
)

Response:

Development Server: Model v2.1 (trained today) Production Server: Model v1.8 (deployed last week)

The development model is 3 versions ahead of production.

AI Assistant Integration

Google ADK Integration

The CMF MCP Server integrates seamlessly with Google's Agent Development Kit (ADK) for building custom agents.

Agent Configuration:

from google import genai
from google.genai.types import Tool, GenerateContentConfig, GoogleSearch

# Configure MCP Toolset
mcp_tool = Tool(
    google_search=GoogleSearch(
        mcp_servers=[
            {
                "name": "cmf-mcp-server",
                "url": "http://localhost:8000/sse"
            }
        ]
    )
)

# Create agent with CMF tools
agent = genai.Client(
    vertexai=True,
    project=PROJECT_ID,
    location=LOCATION
)

# Configure with CMF MCP tools
config = GenerateContentConfig(
    tools=[mcp_tool],
    system_instruction="You are a metadata assistant. Use CMF tools to help users explore ML pipeline metadata."
)

Example Interaction:

# User asks about pipelines
response = agent.models.generate_content(
    model="gemini-2.0-flash-exp",
    contents="What pipelines are in CMF?",
    config=config
)

The agent automatically: 1. Recognizes it needs CMF data 2. Calls the cmf_show_pipelines tool via MCP 3. Formats and presents the results

Claude Desktop Integration

Configure Claude Desktop to use CMF tools for metadata queries.

Configuration (~/.claude_desktop_config.json):

{
  "mcpServers": {
    "cmf-mcp": {
      "transport": "streamable-http",
      "url": "http://localhost:8382/mcp"
    }
  }
}

Example Conversation:

User: What's the lineage for execution abc123 in the Test-env pipeline?

Claude: I'll check the execution lineage for you.

[Uses cmf_execution_lineage tool]

The execution abc123 is part of a training workflow: - It receives processed data from the preprocessing step (ghi789) - Its outputs feed into the evaluation step (def456) - The evaluation results are used by the deployment step (jkl012)

GitHub Copilot Integration

Enable Copilot to query CMF metadata directly from your IDE.

Configuration (.vscode/mcp.json):

{
  "servers": {
    "cmf-mcp-server": {
      "type": "http",
      "url": "http://localhost:8382/mcp"
    }
  }
}

Example Usage in VSCode:

You can ask Copilot directly in your code:

# Ask Copilot: "What artifacts are in the Test-env pipeline?"
# Copilot will use CMF tools to fetch and display the information

Copilot can help you: - Discover available pipelines before writing metadata logging code - Verify artifact types match your pipeline configuration - Check execution history while debugging

Cursor IDE Integration

Cursor provides native MCP support with an intuitive configuration interface.

Setup Steps:

1. Open Cursor Settings → Features → MCP Servers
2. Add CMF MCP Server:
3. Name: cmf-mcp-server
4. Type: http
5. URL: http://localhost:8382/mcp

Example Usage:

Use Cursor's AI chat to query CMF:

User: @cmf-mcp-server What models exist in the training-pipeline?

Cursor AI: [Queries CMF] There are 3 Model artifacts in training-pipeline: - baseline_model.pkl (ID: 10) - improved_model.pkl (ID: 23) - production_model.pkl (ID: 42)

Advanced Use Cases

Pipeline Comparison

Query:

Compare execution times between the last 3 runs of the training pipeline

AI Assistant Workflow:

1. Retrieves executions with cmf_show_executions
2. Extracts timing information
3. Calculates and presents comparison

Response:

Execution Time Comparison: - Run 1 (Feb 7): 2.3 hours - Run 2 (Feb 8): 2.1 hours - Run 3 (Feb 9): 1.9 hours

Execution time has improved by 17% over the last 3 runs.

Artifact Provenance

Query:

Which dataset was used to train the production model?

AI Assistant Workflow:

1. Finds the production model using cmf_show_artifacts
2. Retrieves model card with cmf_show_model_card
3. Examines input artifacts section

Response:

The production model (ID: 42) was trained using: - Primary Dataset: production_training_v2.csv (100K samples) - Validation Dataset: production_validation_v2.csv (20K samples) - Training Date: 2025-02-01

Model Governance Queries

Query:

List all models trained in the last week with their accuracy metrics

AI Assistant Workflow:

1. Retrieves all model artifacts
2. Filters by creation date
3. Fetches model cards for each
4. Extracts accuracy metrics

Response:

Models trained this week:

Model	Training Date	Accuracy
model_v2.1	Feb 9	0.95
model_v2.0	Feb 7	0.93
baseline_v3	Feb 5	0.89

Best Practices

Efficient Querying

1. Start broad, then narrow:

   What pipelines exist? → What executions are in X? → Show lineage for Y
   

2. Validate before querying:

3. Check pipeline names with cmf_show_pipelines before querying executions

4. Verify artifact types with cmf_show_artifact_types before querying artifacts

5. Use specific servers when possible:

6. Include cmfClient_instances parameter when you know the target server
7. Reduces query time and response size

Natural Language Tips

AI assistants work best with clear, specific queries:

Good: - ✅ "What is the execution lineage for UUID abc123 in Test-env?" - ✅ "Show me all Model artifacts in the training-pipeline" - ✅ "Get the model card for model ID 42"

Less Effective: - ❌ "Tell me about the thing" (too vague) - ❌ "Show me everything" (too broad) - ❌ "What happened?" (lacks context)

Error Recovery

If a query fails, AI assistants can usually recover:

Example:

User: Show me executions for pipeline XYZ
AI: The pipeline 'XYZ' was not found. Let me check what pipelines are available.
AI: [Uses cmf_show_pipelines]
AI: I found these pipelines: Test-env, training-pipeline, inference-pipeline.
    Did you mean one of these?

Next Steps

• Tools Reference - Detailed documentation for each tool
• Configuration - Set up multi-server environments
• Quick Start - Get started with the MCP Server