Skip to content

sktime/sktime-mcp

BranchesTags

Repository files navigation

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.

🛠️ Installation

Zero-install via uvx (recommended)

If you have uv installed, no separate installation step is needed. Just update your MCP client config (see Connecting from an LLM Client below) and uvx will handle the rest automatically.

# Verify uv is available
uvx sktime-mcp --help

pip

pip install sktime-mcp

# With optional extras (SQL, forecasting models, file formats)
pip install "sktime-mcp[all]"

Development installation

git clone https://github.com/sktime/sktime-mcp
cd sktime-mcp
python3 -m 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

macOS / Unix-like first-time setup

For macOS or Unix-like shells, create an isolated virtual environment before installing the package:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install sktime-mcp

For development (if you want to modify the source):

python -m pip install -e ".[dev]"

Verify that the MCP server starts:

sktime-mcp

If the sktime-mcp console command is not found (e.g. the script was not placed on your PATH), use the module fallback instead — this is also the recommended form when an MCP client needs to target a specific Python environment:

python -m sktime_mcp.server

Common first-time issues:

Symptom Likely cause Fix
command not found: sktime-mcp Scripts directory not on PATH Run python -m sktime_mcp.server or add .venv/bin to your PATH
ModuleNotFoundError: sktime_mcp Package not installed in the active environment Confirm .venv is active (which python) and re-run pip install sktime-mcp
pip: command not found System pip not available Use python -m pip instead of bare pip
Wrong Python version selected Multiple Python installations Invoke python3 -m venv .venv explicitly and always use python inside the activated environment

Minimal Setup Check

After completing the steps above, confirm the server starts with sktime-mcp. See the macOS / Unix-like first-time setup section for the fallback command and common error solutions.

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.

🚀 Quick Start

Running the MCP Server

Standard Stdio Mode (for MCP Clients)

sktime-mcp

HTTP/SSE Mode via FastAPI (for Web Browsers or ChatGPT)

To expose the MCP server as a REST API over SSE (Server-Sent Events) for direct consumption:

PYTHONPATH=src .venv/bin/uvicorn sktime_mcp.app:app --host 127.0.0.1 --port 8001

This exposes standard SSE on /sse and message passing on /messages/.

Note for ChatGPT Web Users: ChatGPT runs in the cloud and cannot connect to http://127.0.0.1 (you will get an "Unsafe URL" error). You must expose your local server to the internet using a secure tunnel like ngrok:

ngrok http 8001

Then use the provided https://<your-ngrok-id>.ngrok-free.app/sse URL in ChatGPT.

Configuration (Environment Variables)

You can configure the server's behavior at runtime using environment variables:

  • SKTIME_MCP_MAX_RESPONSE_TOKENS: Maximum tokens allowed per tool response (e.g., 10000). If a response exceeds this limit, it is truncated and appended with a notice. Set to 0 (default) for unlimited.
  • SKTIME_MCP_LOG_LEVEL: Server logging verbosity level (DEBUG, INFO, WARNING, ERROR). Defaults to WARNING.
  • SKTIME_MCP_AUTO_FORMAT: Enables or disables automatic time-series formatting during data loading.
  • SKTIME_MCP_JOB_MAX_AGE_HOURS: Maximum hours before completed background jobs are automatically pruned. Defaults to 24.

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

With uvx (recommended — no prior install needed):

{
  "mcpServers": {
    "sktime": {
      "command": "uvx",
      "args": ["sktime-mcp"]
    }
  }
}

With optional extras:

{
  "mcpServers": {
    "sktime": {
      "command": "uvx",
      "args": ["sktime-mcp[forecasting,sql]"]
    }
  }
}

With pip-installed package:

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

⚙️ Configuration

The server can be configured via environment variables:

Environment Variable Description Default
SKTIME_MCP_LOG_LEVEL Logging verbosity (e.g. INFO, DEBUG, WARNING) "WARNING"
SKTIME_MCP_LOG_PATH Optional file path to output logs to in addition to stderr (None)
SKTIME_MCP_AUTO_FORMAT Automatically format time series data on load (true/false) "true"
SKTIME_MCP_JOB_MAX_AGE_HOURS Maximum age in hours before background jobs are cleared 24
SKTIME_MCP_JOB_CLEANUP_INTERVAL Interval in seconds for periodic job cleanup checks 3600

📚 Available Tools

The sktime-mcp server exposes a rich suite of tools categorized logically. Every tool is fully typed, specifies clear return schemas, and aligns with native sktime naming conventions wherever possible.

Category 1: Discovery & Registry Tools

These tools enable the LLM to inspect the native sktime registry, search for estimators matching specific criteria, and understand their capability profiles.

1. list_estimators

Discover estimator classes from the sktime registry by task type, capabilities, or name query.

  • Arguments:
    • task (str, optional): Filter by task type. Valid values: "forecasting", "classification", "regression", "transformation", "clustering", "detection".
    • tags (dict[str, Any], optional): Key-value pairs filtering by capability tags (e.g., {"capability:pred_int": true, "handles-missing-data": true}).
    • query (str, optional): Case-insensitive substring search on the estimator's class name or docstring.
    • limit (int, optional, default=50): Maximum number of results to return.
    • offset (int, optional, default=0): Number of entries to skip for pagination.
  • Returns: 
    {
  "success": true,
  "estimators": [
    {
      "name": "ARIMA",
      "task": "forecasting",
      "description": "Autoregressive Integrated Moving Average (ARIMA) model...",
      "tags": {
        "capability:pred_int": true,
        "handles-missing-data": false
      },
      "import_path": "sktime.forecasting.arima.ARIMA"
    }
  ],
  "total": 1
}

2. describe_estimator

Get full documentation, hyperparameters, capability tags, and Python import path for a specific named estimator class.

  • Arguments:
    • estimator (str, required): Name of the estimator class (e.g., "ARIMA", "NaiveForecaster").
  • Returns: 
    {
  "success": true,
  "name": "ARIMA",
  "description": "Autoregressive Integrated Moving Average...",
  "params": {
    "order": {
      "type": "tuple",
      "default": [1, 0, 0],
      "description": "The (p,d,q) order of the model."
    }
  },
  "tags": {
    "capability:pred_int": true
  },
  "import_path": "sktime.forecasting.arima.ARIMA"
}

3. get_available_tags

List all queryable capability tags across the sktime registry with their descriptions and expected value types.

  • Arguments: None.
  • Returns: 
    {
  "success": true,
  "tags": {
    "capability:pred_int": {
      "type": "bool",
      "description": "Whether the forecaster can compute prediction intervals.",
      "scitype": "forecaster"
    }
  }
}

Category 2: Instantiation & Composition Tools

These tools manage the creation and lifecycle of model instances in server memory.

4. instantiate

Unify single estimators and multi-step composite pipelines under a single instantiator, returning an active estimator_handle.

  • Arguments:
    • estimator (str | list[str], required): Either a single estimator class name or an ordered list of class names representing a pipeline (e.g., ["Detrender", "ARIMA"]).
    • params (dict[str, Any] | list[dict[str, Any]], optional): Hyperparameters for the single estimator, or a list of parameter dicts corresponding element-wise to the pipeline steps.
  • Returns: 
    {
  "success": true,
  "estimator_handle": "est_abc123",
  "structure": "Detrender ➔ ARIMA",
  "params": [{}, {"order": [1, 1, 1]}]
}

5. validate_pipeline

Verify the compatibility and validity of a list of estimator steps in pipeline order before instantiation.

  • Arguments:
    • components (list[str], required): List of estimator class names in pipeline order.
  • Returns: 
    {
  "success": true,
  "valid": true,
  "errors": [],
  "warnings": []
}

6. list_handles

List all active estimator and pipeline handles currently allocated in server memory.

  • Arguments: None.
  • Returns: 
    {
  "success": true,
  "handles": [
    {
      "handle": "est_abc123",
      "class_name": "Pipeline",
      "is_fitted": false,
      "created_at": "2026-05-19T10:30:00Z"
    }
  ]
}

7. release_handle

Free an estimator or pipeline handle from server memory to prevent leaks.

  • Arguments:
    • handle (str, required): The handle ID to release.
  • Returns: 
    {
  "success": true,
  "message": "Handle 'est_abc123' released successfully."
}

Category 3: Execution & Application Tools

Execution tools apply estimator handles to data. Rather than segregating custom files and built-in demos, arguments like y and X accept either a registered data_handle or a built-in demo dataset name (e.g., "airline").

8. fit / fit_async

Fit an instantiated estimator or pipeline on a target time series y and optional exogenous data X.

  • Arguments:
    • estimator_handle (str, required): Active estimator handle.
    • y (str, required): Target time series. Can be an active data_handle or a built-in demo dataset name (e.g., "airline").
    • X (str, optional): Exogenous time series (active data_handle or demo dataset name).
    • fit_params (dict[str, Any], optional): Extra parameters forwarded to the underlying fit call (e.g., fh if required at fit time).
  • Returns (fit): 
    {
  "success": true,
  "estimator_handle": "est_abc123",
  "message": "Estimator fitted successfully."
}
  • Returns (fit_async): 
    {
  "success": true,
  "job_id": "job_fit_xyz789",
  "message": "Fitting job started in background."
}

9. predict / predict_async

Generate predictions from a fitted estimator or pipeline.

  • Arguments:
    • estimator_handle (str, required): Fitted estimator handle.
    • X (str, optional): Exogenous/feature series (active data_handle or demo name).
    • fh (list[int] | list[str] | int, optional): Forecast horizon (for forecasters).
    • predict_params (dict[str, Any], optional): Extra parameters forwarded to predict.
  • Returns (predict): 
    {
  "success": true,
  "predictions": [
    {"index": "1961-01", "value": 450.2},
    {"index": "1961-02", "value": 455.1}
  ],
  "fh": [1, 2]
}
  • Returns (predict_async): 
    {
  "success": true,
  "job_id": "job_pred_123456",
  "message": "Prediction job started in background."
}

10. predict_interval

Generate prediction intervals from a fitted probabilistic forecaster.

  • Arguments:
    • estimator_handle (str, required): Fitted estimator handle.
    • X (str, optional): Exogenous/feature series.
    • fh (list[int] | list[str] | int, optional): Forecast horizon.
    • coverage (list[float] | float, optional, default=0.9): Nominal coverage interval(s) in (0, 1).
  • Returns: 
    {
  "success": true,
  "intervals": [
    {
      "index": "1961-01",
      "lower_0.9": 435.1,
      "upper_0.9": 465.3
    }
  ],
  "coverage": [0.9]
}

11. predict_quantiles

Generate quantile forecasts from a fitted probabilistic forecaster.

  • Arguments:
    • estimator_handle (str, required): Fitted estimator handle.
    • X (str, optional): Exogenous series.
    • fh (list[int] | list[str] | int, optional): Forecast horizon.
    • alpha (list[float] | float, required): Quantile values to compute in (0, 1).
  • Returns: 
    {
  "success": true,
  "quantiles": [
    {
      "index": "1961-01",
      "quantile_0.05": 420.3,
      "quantile_0.95": 480.1
    }
  ],
  "alpha": [0.05, 0.95]
}

12. transform

Apply a transformer to data, registering the transformed output as a new data_handle.

  • Arguments:
    • estimator_handle (str, required): Instantiated transformer handle.
    • X (str, required): Series/DataFrame to transform (active data_handle or demo name).
    • y (str, optional): Target series (active data_handle or demo name).
    • action (str, optional, default="transform"): Action to perform. Valid values: "transform", "fit_transform", "inverse_transform".
  • Returns: 
    {
  "success": true,
  "data_handle": "data_trans_999",
  "metadata": {
    "shape": [144, 1],
    "frequency": "M",
    "columns": ["value_transformed"]
  }
}

13. update

Update a fitted forecaster/model with new incoming streaming data (online learning).

  • Arguments:
    • estimator_handle (str, required): Fitted estimator handle.
    • y (str, required): New target observations (active data_handle or demo name).
    • X (str, optional): New exogenous observations (active data_handle or demo name).
    • update_params (dict[str, Any], optional): Extra parameters passed to the update call.
  • Returns: 
    {
  "success": true,
  "message": "Model updated successfully with 12 new observations."
}

14. get_fitted_params

Retrieve learnable coefficients, components, or parameters from a fitted estimator.

  • Arguments:
    • estimator_handle (str, required): Fitted estimator handle.
  • Returns: 
    {
  "success": true,
  "fitted_params": {
    "cutoff": "1960-12",
    "sigma2": 0.0431,
    "ar_coefficients": [0.824]
  }
}

Category 4: Evaluation Tools

These tools streamline model validation and scoring pipelines.

15. evaluate / evaluate_async

Cross-validate an estimator on a dataset to assess performance.

  • Arguments:
    • estimator_handle (str, required): Instantiated estimator handle.
    • y (str, required): Target time series (active data_handle or demo name).
    • X (str, optional): Exogenous time series.
    • cv_folds (int, optional, default=3): Number of folds for cross-validation.
    • metric (str, optional): Performance metric name (e.g., "MeanAbsolutePercentageError").
  • Returns (evaluate): 
    {
  "success": true,
  "metrics": {
    "mean_absolute_percentage_error": 0.054
  },
  "fold_results": [
    {"fold": 0, "score": 0.049},
    {"fold": 1, "score": 0.058}
  ]
}
  • Returns (evaluate_async): 
    {
  "success": true,
  "job_id": "job_eval_111",
  "message": "Evaluation job started in background."
}

Category 5: Data Management Tools

These tools manage input and output datasets, providing caching, standardization, and memory reclamation.

16. list_available_data

List all built-in demo datasets and active user-loaded data handles in a single response.

  • Arguments:
    • is_demo (bool, optional): Filter to built-in system demos (true) or active in-memory handles (false). Omit to get both.
  • Returns: 
    {
  "success": true,
  "system_demos": ["airline", "sunspots", "lynx"],
  "active_handles": [
    {
      "data_handle": "data_abc123",
      "shape": [144, 2],
      "columns": ["target", "exogenous"],
      "frequency": "M"
    }
  ]
}

17. load_data_source / load_data_source_async

Load custom files, SQL database queries, URLs, or inline JSON into the server as a data_handle.

  • Arguments:
    • config (dict[str, Any], required): Source configuration dictionary. Must contain:
      • "type" (str, required): "pandas" (inline dict), "file" (CSV/Excel/Parquet), "sql" (DB connection), or "url".
      • Additional keys based on "type" (e.g., "path", "url", "connection_string", "query", "time_column", "target_column").
  • Returns (load_data_source): 
    {
  "success": true,
  "data_handle": "data_abc123",
  "metadata": {
    "shape": [500, 1],
    "columns": ["sales"],
    "frequency": "D"
  }
}
  • Returns (load_data_source_async): 
    {
  "success": true,
  "job_id": "job_load_444",
  "message": "Asynchronous loading started."
}

18. save_data

Persist an in-memory data_handle (such as predictions or transformed series) back to disk.

  • Arguments:
    • data_handle (str, required): In-memory data handle to save.
    • path (str, required): Local filesystem path where the file will be saved.
    • format (str, optional): Output format (inferred from path extension if omitted: "csv", "parquet", "json").
  • Returns: 
    {
  "success": true,
  "saved_path": "/home/user/forecasts.csv",
  "message": "Data saved successfully."
}

19. format_time_series

Clean, fill missing values, deduplicate, and standardize loaded time series data.

  • Arguments:
    • data_handle (str, required): Target data handle.
    • auto_infer_freq (bool, optional, default=true): Re-infer time delta frequency.
    • fill_missing (bool, optional, default=true): Interpolate missing values using forward/backward fills.
    • remove_duplicates (bool, optional, default=true): Deduplicate timestamps.
  • Returns: 
    {
  "success": true,
  "data_handle": "data_abc123",
  "changes_applied": ["inferred frequency: M", "filled 3 missing values"]
}

20. release_data_handle

Free a data handle and its contents from server memory.

  • Arguments:
    • data_handle (str, required): Handle ID to release.
  • Returns: 
    {
  "success": true,
  "message": "Data handle 'data_abc123' freed from memory."
}

Category 6: Persistence & Code Generation

These tools manage the serialization of estimator instances and generation of production-ready source code.

21. save_model

Serialize an estimator blueprint or fitted model handle to disk using sktime-MLflow integration.

  • Arguments:
    • estimator_handle (str, required): Estimator or pipeline handle to save.
    • path (str, required): Filesystem path or MLflow artifact URI.
    • mlflow_params (dict[str, Any], optional): Extra parameters for MLflow.
  • Returns: 
    {
  "success": true,
  "saved_path": "/path/to/model_dir",
  "is_fitted": true,
  "message": "Fitted model saved successfully."
}

22. load_model

Reload a serialized blueprint or fitted model back into an active estimator_handle.

  • Arguments:
    • path (str, required): Filesystem path to the model directory.
  • Returns: 
    {
  "success": true,
  "estimator_handle": "est_rehydrated_999",
  "is_fitted": true
}

23. export_code

Generate standalone, executable Python code to reproduce an estimator's structure and execution.

  • Arguments:
    • handle (str, required): Handle ID of the estimator/pipeline.
    • var_name (str, optional, default="model"): Variable name in the generated code block.
    • include_fit_example (bool, optional, default=false): Whether to inject boilerplate loading/fitting logic.
    • y (str, optional): Dataset name for the fitting logic (default: "airline").
  • Returns: 
    {
  "success": true,
  "code": "from sktime.forecasting.arima import ARIMA

model = ARIMA(order=(1, 1, 1)) ..." }


---

### Category 7: Background Jobs

These tools poll, track, and manage asynchronous task executions.

#### 24. `check_job_status`
Inspect the status, execution progress, and results of a background job.
* **Arguments:**
* `job_id` (`str`, required): Unique ID of the job.
* **Returns:**
```json
{
  "success": true,
  "job_id": "job_fit_xyz789",
  "status": "COMPLETED",
  "progress_pct": 100.0,
  "error": null,
  "result": {
    "estimator_handle": "est_fitted_123",
    "message": "Estimator fitted successfully."
  }
}

25. list_jobs

List all active and historical background jobs.

  • Arguments:
    • status (str, optional): Filter by status ("pending", "running", "completed", "failed", "cancelled").
    • limit (int, optional, default=20): Pagination limit.
  • Returns: 
    {
  "success": true,
  "jobs": [
    {
      "job_id": "job_fit_xyz789",
      "status": "completed",
      "created_at": "2026-05-19T10:32:00Z"
    }
  ]
}

26. cancel_job

Prune or terminate an active background job.

  • Arguments:
    • job_id (str, required): Unique ID of the job.
    • delete (bool, optional, default=false): If true, deletes the job tracking record from memory.
  • Returns: 
    {
  "success": true,
  "message": "Job 'job_fit_xyz789' terminated."
}

🔄 Example LLM Flows

Below are examples demonstrating how an LLM utilizes these redesigned tools to construct, validate, and execute workflows.

Flow 1: Simple Probabilistic Forecasting

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

LLM Steps:

  1. Discover Models The LLM queries for forecasting models supporting prediction intervals.

    // list_estimators
{
  "task": "forecasting",
  "tags": {
    "capability:pred_int": true
  }
}

    Returns: A list containing "ARIMA" and "NaiveForecaster".

  2. Inspect Choice The LLM inspects the parameter schema of "ARIMA".

    // describe_estimator
{
  "estimator": "ARIMA"
}

  3. Instantiate The LLM instantiates the ARIMA model with specific parameters.

    // instantiate
{
  "estimator": "ARIMA",
  "params": {
    "order": [1, 1, 1]
  }
}

    Returns: {"success": true, "estimator_handle": "est_arima_456", ...}

  4. Fit the Model The LLM fits the model to the built-in "airline" demo dataset.

    // fit
{
  "estimator_handle": "est_arima_456",
  "y": "airline"
}

    Returns: {"success": true, "estimator_handle": "est_arima_456", ...}

  5. Generate Forecast Intervals The LLM computes a 90% confidence prediction interval for the next 12 months.

    // predict_interval
{
  "estimator_handle": "est_arima_456",
  "fh": 12,
  "coverage": 0.90
}

    Returns: Predicton interval upper and lower bounds.

Flow 2: Pipeline Composition & Forecasting

User Prompt: "Compose a forecasting pipeline that includes seasonal deseasonalization, detrending, and ARIMA."

LLM Steps:

  1. Validate Composition The LLM validates compatibility of the steps before allocating handles.

    // validate_pipeline
{
  "components": ["ConditionalDeseasonalizer", "Detrender", "ARIMA"]
}

    Returns: {"success": true, "valid": true}

  2. Instantiate Composite Pipeline The LLM creates the full pipeline in a single call.

    // instantiate
{
  "estimator": ["ConditionalDeseasonalizer", "Detrender", "ARIMA"],
  "params": [{}, {}, {"order": [1, 1, 1]}]
}

    Returns: {"success": true, "estimator_handle": "est_pipe_789", "structure": "ConditionalDeseasonalizer ➔ Detrender ➔ ARIMA"}

  3. Fit the Pipeline

    // fit
{
  "estimator_handle": "est_pipe_789",
  "y": "airline"
}

  4. Predict

    // predict
{
  "estimator_handle": "est_pipe_789",
  "fh": 12
}

    Returns: The predictions generated by the pipeline.

📁 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/                   # Sphinx 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:

ruff format --check .
ruff check .
pytest

Pre-Commit Hooks (Recommended)

To ensure your code meets quality standards before pushing, install the pre-commit hooks:

make install-hooks

This will automatically run Ruff and Pytest on your code every time you make a commit.

About

An MCP (Model Context Protocol) layer that exposes sktime’s native registry and semantics to an LLM

sktime-mcp.readthedocs.io/en/latest/

Topics

sktime mcp-server

Resources

Readme

License

BSD-3-Clause license
Activity
Custom properties

Stars

24 stars

Watchers

1 watching

Forks

111 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages