sktime-mcp

MCP (Model Context Protocol) layer for sktime - Registry-Driven for LLMs

A semantic engine that exposes sktime's native registry and semantics to Large Language Models, enabling them to:

• 🔍 Discover valid estimators
• 🧠 Reason about estimator capabilities
• 🔗 Compose compatible estimators
• ⚡ Execute real sktime workflows on real data

🎯 Design Philosophy

This MCP is not just documentation or static code analysis. It is a semantic engine for programmatic model usage.

Key Principles

1. sktime as Source of Truth - No AST parsing, no repo indexing, no heuristics. All structure comes from all_estimators, estimator tags, and sktime's API contracts.

2. Registry-First - Instead of File → Class → Infer Relationships, we do Registry → Semantics → Safe Execution.

3. Minimal MCP Surface - Exposes only what an LLM needs: Discovery, Description, Instantiation, Execution.

🛠️ Installation

# Install from source
pip install -e .

# With all optional dependencies
pip install -e ".[all]"

# Development installation
pip install -e ".[dev]"

🚀 Quick Start

Running the MCP Server

# Start the MCP server
sktime-mcp

# Or run directly
python -m sktime_mcp.server

Connecting from an LLM Client

The server uses stdio transport by default, compatible with Claude Desktop and other MCP clients.

Add to your Claude Desktop config (~/.config/claude/claude_desktop_config.json):

{
  "mcpServers": {
    "sktime": {
      "command": "sktime-mcp"
    }
  }
}

📚 Available Tools

Discovery & Search

1. list_estimators

Discover estimators by task type and capability tags.

Arguments:

• task (optional): Task type filter ("forecasting", "classification", "regression", "transformation", "clustering")
• tags (optional): Filter by capability tags (e.g., {"capability:pred_int": true})
• limit (optional): Maximum results (default: 50)

Example:

{
  "task": "forecasting",
  "tags": {
    "capability:pred_int": true
  },
  "limit": 10
}

Returns: List of matching estimators with name, task, and summary info.

---

2. search_estimators

Search estimators by name or description using text query.

Arguments:

• query (required): Search string (case-insensitive)
• limit (optional): Maximum results (default: 20)

Example:

{
  "query": "ARIMA",
  "limit": 5
}

Returns: List of estimators matching the search query.

---

3. describe_estimator

Get detailed information about a specific estimator's capabilities.

Arguments:

• estimator (required): Name of the estimator (e.g., "ARIMA", "NaiveForecaster")

Example:

{
  "estimator": "ARIMA"
}

Returns: Full estimator details including tags, hyperparameters, docstring, and module path.

---

4. get_available_tags

List all queryable capability tags across all estimators.

Arguments: None

Returns: List of all available tags (e.g., ["capability:pred_int", "handles-missing-data", ...])

---

Instantiation

5. instantiate_estimator

Create a single estimator instance and return a handle.

Arguments:

• estimator (required): Name of the estimator to instantiate
• params (optional): Hyperparameters for the estimator

Example:

{
  "estimator": "ARIMA",
  "params": {
    "order": [1, 1, 1],
    "suppress_warnings": true
  }
}

Returns: {"success": true, "handle": "est_abc123", "estimator": "ARIMA", "params": {...}}

---

6. instantiate_pipeline

Create a complete pipeline from a list of components (transformers → forecaster).

Arguments:

• components (required): List of estimator names in pipeline order
• params_list (optional): List of parameter dicts for each component

Example:

{
  "components": ["ConditionalDeseasonalizer", "Detrender", "ARIMA"],
  "params_list": [{}, {}, {"order": [1, 1, 1]}]
}

Returns: {"success": true, "handle": "est_xyz789", "pipeline": "ConditionalDeseasonalizer → Detrender → ARIMA", ...}

Note: This solves the "steps problem" - you don't need to instantiate components separately!

---

📖 Documentation

Project documentation lives in docs/ and can be served locally with MkDocs:

pip install -e ".[dev]"
mkdocs serve

The MkDocs config is in mkdocs.yml.

Validation

7. validate_pipeline

Check if a proposed pipeline composition is valid before instantiation.

Arguments:

• components (required): List of estimator names in pipeline order

Example:

{
  "components": ["Detrender", "ARIMA"]
}

Returns: {"valid": true/false, "errors": [...], "warnings": [...], "suggestions": [...]}

---

Execution

8. fit_predict

Execute a complete workflow: load dataset, fit estimator, and generate predictions.

Arguments:

• estimator_handle (required): Handle from instantiate_estimator or instantiate_pipeline
• dataset (required): Dataset name (e.g., "airline", "sunspots", "lynx")
• horizon (optional): Forecast horizon (default: 12)

Example:

{
  "estimator_handle": "est_abc123",
  "dataset": "airline",
  "horizon": 12
}

Returns: {"success": true, "predictions": {1: 450.2, 2: 455.1, ...}, "horizon": 12}

---

Datasets

9. list_datasets

List all available demo datasets for testing and experimentation.

Arguments: None

Returns: {"success": true, "datasets": ["airline", "sunspots", "lynx", "shampoo", ...]}

---

Handle Management

10. list_handles

List all active estimator handles and their status.

Arguments: None

Returns: List of active handles with metadata (estimator name, fitted status, creation time)

---

11. release_handle

Release an estimator handle and free memory.

Arguments:

• handle (required): Handle ID to release

Example:

{
  "handle": "est_abc123"
}

Returns: {"success": true, "message": "Handle released"}

🔄 Example LLM Flows

Flow 1: Simple Forecasting

User Prompt: "Forecast monthly airline passengers using a probabilistic model."

LLM Steps:

1. Discover Models

   list_estimators(task="forecasting", tags={"capability:pred_int": true})
   

2. Inspect Choice

   describe_estimator(estimator="ARIMA")
   

3. Instantiate

   instantiate_estimator(estimator="ARIMA", params={"order": [1,1,1]})
   → Returns: {"handle": "est_abc123"}
   

4. Execute

   fit_predict(estimator_handle="est_abc123", dataset="airline", horizon=12)
   → Returns: {"predictions": {1: 450.2, 2: 455.1, ...}}
   

Flow 2: Pipeline Forecasting ⭐

User Prompt: "Forecast with deseasonalization and detrending preprocessing."

LLM Steps:

1. Validate Composition

   validate_pipeline(components=["ConditionalDeseasonalizer", "Detrender", "ARIMA"])
   → Returns: {"valid": true}
   

2. Instantiate Pipeline (single call!)

   instantiate_pipeline(
     components=["ConditionalDeseasonalizer", "Detrender", "ARIMA"],
     params_list=[{}, {}, {"order": [1,1,1]}]
   )
   → Returns: {"handle": "est_xyz789", "pipeline": "ConditionalDeseasonalizer → Detrender → ARIMA"}
   

3. Execute

   fit_predict(estimator_handle="est_xyz789", dataset="airline", horizon=12)
   → Returns: {"predictions": {...}}
   

📁 Project Structure

sktime_mcp/
├── src/sktime_mcp/
│   ├── server.py           # MCP server entry point
│   ├── registry/           # Registry interface & tag resolver
│   ├── composition/        # Pipeline composition validator
│   ├── runtime/            # Execution engine & handle management
│   └── tools/              # MCP tool implementations
├── examples/               # Usage examples
└── tests/                  # Test suite

🧪 Running Tests

pytest tests/