sktime-mcp

Read the Documentation | PyPI Package

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, and model persistence.

🛠️ Prerequisites

• Python 3.10+ (3.9 is listed in pyproject.toml but the mcp package requires 3.10+)
• pip package manager

🛠️ Installation

Virtual Environment Setup

It is recommended to use a virtual environment:

python3 -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

Package Installation

The recommended way to install is using python3 -m pip:

# Install from source
python3 -m pip install -e .

# With all optional dependencies
python3 -m pip install -e ".[all]"

# Development installation
python3 -m pip install -e ".[dev]"

Alternatively, you can use pip:

# Install from source
pip install -e .

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

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

🧭 Beginner Setup (First‑Time Users)

If you are new to sktime‑mcp or to MCP‑based workflows, this section provides a minimal starting point to help you verify that your setup is working correctly.

What is MCP?

The Model Context Protocol (MCP) allows Large Language Models (LLMs) to discover, reason about, and execute sktime workflows programmatically. This project exposes sktime’s estimator registry and semantics in a structured way so that LLMs can safely compose and run real time‑series pipelines.

Prerequisites

• Python 3.10 or newer
• A working Python virtual environment (recommended)
• pip installed

Minimal Setup Check

After installing the package, you can verify that the MCP server starts correctly by running:

sktime-mcp

Note: On Windows, the sktime-mcp command may be installed to a directory not on your PATH (e.g., %APPDATA%\Python\Python3xx\Scripts). Either add that directory to your PATH or use python -m sktime_mcp.server instead.

Note: On some systems (like macOS), pip may not be available in the path. In such cases, use python3 -m pip to ensure the command runs with the intended Python version.

🚀 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, Claude Code, and other MCP clients.

Claude Desktop — add to your config file:

Platform	Config path
macOS	~/Library/Application Support/Claude/claude_desktop_config.json
Linux	~/.config/claude/claude_desktop_config.json
Windows	%APPDATA%\Claude\claude_desktop_config.json

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

If sktime-mcp is not on your PATH, use the full path to the executable or use python -m sktime_mcp.server as the command instead.

📚 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}

---

9. save_model

Persist a fitted estimator or pipeline handle to a local filesystem path using sktime.utils.mlflow_sktime.save_model.

Arguments:

• estimator_handle (required): Handle from instantiate_estimator or instantiate_pipeline
• path (required): Local filesystem path where the model should be saved
• mlflow_params (optional): Extra keyword arguments forwarded to sktime.utils.mlflow_sktime.save_model

Example:

{
  "estimator_handle": "est_abc123",
  "path": "/absolute/path/to/model_dir",
  "mlflow_params": {
    "serialization_format": "cloudpickle"
  }
}

Returns: {"success": true, "saved_path": "/absolute/path/to/model_dir", "message": "Model saved successfully to '/absolute/path/to/model_dir'"}

Note: This tool requires MLflow to be available in the server environment.

---

Data Availability

10. list_available_data

List all available data — both system demo datasets and active user-loaded data handles — in a single unified response.

Arguments:

• is_demo (optional): true returns only demo datasets, false returns only active data handles, omit to get both

Returns: {"success": true, "system_demos": ["airline", "sunspots", ...], "active_handles": [...], "total": 8}

---

Data Loading

11. load_data_source

Load data from various sources (CSV/Parquet files, pandas DataFrames, SQL databases, URLs).

Arguments:

• config (required): Data source configuration object. Must include "type" ("pandas", "sql", "file", "url").

Example:

{
  "config": {
    "type": "file",
    "path": "/absolute/path/to/data.csv",
    "time_column": "date",
    "target_column": "value"
  }
}

Returns: {"success": true, "data_handle": "data_abc123", "metadata": {...}}

---

12. fit_predict_with_data

Fit an estimator and generate predictions using a custom data handle (instead of a demo dataset).

Arguments:

• estimator_handle (required): Handle from instantiate_estimator
• data_handle (required): Handle from load_data_source
• horizon (optional): Forecast horizon (default: 12)

Returns: Same format as fit_predict.

---

Code & Model Export

13. export_code

Export an estimator or pipeline as executable Python code.

Arguments:

• handle (required): Handle ID of the estimator/pipeline
• var_name (optional): Variable name in generated code (default: "model")
• include_fit_example (optional): Include a fit/predict example (default: false)

Returns: {"success": true, "code": "from sktime..."}

---

Background Jobs

14. fit_predict_async

Non-blocking version of fit_predict. Returns a job_id immediately; use check_job_status to poll.

Arguments: Same as fit_predict.

Returns: {"success": true, "job_id": "abc-123", "message": "Training job started..."}

---

15. check_job_status

Check the status and progress of a background job.

Arguments:

• job_id (required): Job ID to check

---

16. list_jobs

List all background jobs with optional status filter.

Arguments:

• status (optional): Filter by status ("pending", "running", "completed", "failed", "cancelled")
• limit (optional): Maximum results (default: 20)

---

Data Formatting

17. format_time_series

Automatically format time series data (infer frequency, remove duplicates, fill missing values).

Arguments:

• data_handle (required): Handle from load_data_source
• auto_infer_freq (optional): Infer and set frequency (default: true)
• fill_missing (optional): Fill missing values (default: true)
• remove_duplicates (optional): Remove duplicate timestamps (default: true)

🔄 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 & job management
│   ├── data/               # Data adapters (file, pandas, SQL, URL)
│   └── tools/              # MCP tool implementations
├── docs/                   # MkDocs documentation source
├── examples/               # Usage examples
└── tests/                  # Test suite

🧪 Running Tests

pytest tests/

Local Quality Checks

Run standardized local checks before raising a PR:

make check

Auto-fix formatting and fixable lint issues:

make format-fix

If make is unavailable (common on Windows), run the equivalent commands:

python -m black --check .
python -m ruff check .
python -m pytest