data-science-mcp

data-science-mcp is a Model Context Protocol (MCP) server for reproducible data-science workflows. Its current domain is time-series analysis and forecasting: it exposes a structured tool catalogue for loading tabular time-series data, repairing data-quality issues, running diagnostics, and producing forecasts with statistical, foundation-model, and AutoML methods. The server is organized so additional data-science domains can be added as new tool groups over time.

The server was developed as a scientific-paper implementation artifact. This README is written so it can be used as an appendix describing the implemented MCP server, its public interface, and representative forecasting tasks.

Project Metadata

• Author: LiChenStuttgart, 3380836@gmail.com

• Package name: data-science-mcp

• Current package version: 0.1.0

• License: MIT

• Python requirement: >=3.10

• MCP transport: stdio through mcp.server.fastmcp.FastMCP

Related MCP server: forecast-mcp

Overall Summary

The MCP server acts as a bridge between an LLM-capable MCP client and a time-series analysis library. It registers 12 public tools and 2 MCP resources. The main workflow is:

1. Discover tools with list_tools or the guide://data-science resource.

2. Load tabular data into the canonical SeriesCollection contract with time_series_loader.

3. Handle missing values and outliers while preserving audit metadata.

4. Run stationarity, seasonality, or structural-break diagnostics.

5. Forecast with ARIMA/SARIMA, Holt-Winters exponential smoothing, Chronos-2, Toto 2.0, or AutoGluon TimeSeries.

Downstream tools accept one of three input styles:

• series_collection: the dataset object returned by time_series_loader or by a previous quality-handling tool.

• data: a direct list of numeric values, which is converted into a synthetic single-series collection.

• file_path plus loader arguments such as time_column, value_columns, dimension_columns, and frequency.

Tool outputs are JSON strings by default. Most tools also accept output_format as json, markdown, or text, and save_path to persist a rendered result.

MCP Resources

• guide://data-science: live registry-generated tool catalogue, grouped by server, data preparation, data quality, diagnostics, and forecasting.

• guide://data-contracts: canonical input and output schema documentation for SeriesCollection, quality flags, and forecast result objects.

Tool Catalogue

Data Contract Notes

time_series_loader returns a top-level object with ok: true and dataset. The dataset is a SeriesCollection with:

• source: file path, file type, and load timestamp.

• schema: time column, value columns, dimension columns, normalized frequency, requested frequency, and aggregation.

• time_index: regular timestamps generated from the requested frequency.

• series: one item per metric and dimension combination.

• quality: row counts, series count, timestamp count, and warnings.

Accepted frequency aliases include daily (day, daily, D), monthly (month, monthly, M), quarterly (quarter, quarterly, Q), and yearly (year, annual, yearly, Y). They are normalized to D, M, Q, or Y.

For wide scientific tables where columns are separate metrics rather than dimensions, pass dimension_columns: [] explicitly. If dimension_columns is omitted, the loader infers non-time, non-value columns as dimensions.

Quality handlers preserve audit fields in series[].metadata, including imputation_flags, imputation_strategy, outlier_flags, outlier_detection_method, and outlier_handling_method.

Running The Server

Create or activate a Python environment, then install the package. The scientific extra is recommended for full local data loading and diagnostics:

python -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[scientific]'
python -m data_science_mcp

The console script is also available after installation:

data-science-mcp --version
data-science-mcp

Example MCP client configuration:

{
  "mcpServers": {
    "data-science-mcp": {
      "command": "/Users/lchen/github_repos/data-science-mcp/.venv/bin/python",
      "args": ["-m", "data_science_mcp"]
    }
  }
}

Use an absolute virtualenv Python path for MCP clients. Most clients launch servers without activating the user's shell environment, so python -m data_science_mcp can fail outside this repository if the package is not installed in that interpreter.

Optional Runtime Dependencies

Some tools are always registered but require optional runtime dependencies. If a dependency is missing, the tool returns a structured tool_unavailable or dependency error payload.

Chronos-2

time_series_chronos2_forecast uses Amazon Chronos-2 for zero-shot point and quantile forecasts.

python -m pip install -e '.[chronos2]'
python scripts/manage_modelweights.py download
python scripts/manage_modelweights.py verify

The default amazon/chronos-2 ID resolves to modelweights/chronos-2. Runtime inference does not automatically download the canonical model. Set TS_MCP_MODEL_DEVICE to cpu, cuda, or another supported device value. Set TS_MCP_MODEL_WEIGHTS_DIR to use a different local model root.

Toto 2.0

time_series_toto2_forecast uses Datadog Toto 2.0 for zero-shot fixed-quantile forecasts and multivariate grouped inference.

python -m pip install -e '.[toto2]'
python scripts/manage_modelweights.py download
python scripts/manage_modelweights.py verify

Toto 2.0 requires Python 3.12 or newer. The default Datadog/Toto-2.0-22m ID resolves to modelweights/toto-2.0-22m. Supported quantile outputs are fixed levels from 0.1 through 0.9, and 0.5 must be included because it is used as the point forecast.

AutoML

time_series_automl_forecast uses AutoGluon TimeSeries:

python -m pip install -e '.[automl]'

Install this extra into the same Python environment used by the MCP server. AutoML accepts past-only covariates, known-future covariates, generated calendar covariates, and optional model lists such as SeasonalNaive, ETS, or Chronos2. Temporary AutoGluon artifacts are cleaned up unless model_path is supplied.

Example Forecasting Tasks

The following examples are phrased as user prompts for an MCP client. The JSON snippets show representative tool calls and abbreviated possible results from the current implementation. Numeric values can vary across dependency versions, random seeds, optional model weights, and hardware.

Example 1: Monthly car registrations with seasonal ARIMA

Prompt:

Load docs/First registration of brand new passenger cars - selected4SAC_columns.csv
as monthly time-series data. The first column is the month, and Belgium,
France, Germany, Greece, and United Kingdom are separate value series. Forecast
Germany for the next 3 months with automatic seasonal ARIMA and report MAPE.

Tool calls:

{
  "name": "time_series_loader",
  "arguments": {
    "file_path": "docs/First registration of brand new passenger cars - selected4SAC_columns.csv",
    "time_column": "Country",
    "value_columns": ["Belgium", "France", "Germany", "Greece", "United Kingdom"],
    "dimension_columns": [],
    "frequency": "month"
  }
}

{
  "name": "time_series_arima",
  "arguments": {
    "series_collection": "<dataset from time_series_loader>",
    "target_metrics": ["Germany"],
    "auto_order": true,
    "max_p": 2,
    "max_d": 1,
    "max_q": 2,
    "max_seasonal_p": 1,
    "max_seasonal_d": 1,
    "max_seasonal_q": 1,
    "seasonal_period": 12,
    "forecast_steps": 3,
    "holdout_size": 12
  }
}

Possible result excerpt:

{
  "ok": true,
  "forecast_results": [
    {
      "series_id": "Germany",
      "model": "ARIMA",
      "order": [0, 1, 2],
      "seasonal_order": [0, 1, 1, 12],
      "forecast_index": ["2025-01-31", "2025-02-28", "2025-03-31"],
      "forecast": [199588.43, 214373.91, 278359.04],
      "metrics": {
        "holdout_mape": 10.5686,
        "in_sample_mape": 8.6666
      }
    }
  ]
}

Example 2: Yearly electric-car stock with ARIMA and Holt-Winters

Prompt:

Load docs/Yearly Global Electric Car Stock.csv. Forecast Total (millions) for
the next 3 years using ARIMA(1,1,1) and Holt-Winters exponential smoothing.
Use a 3-observation holdout and compare the validation MAPE.

Tool calls:

{
  "name": "time_series_loader",
  "arguments": {
    "file_path": "docs/Yearly Global Electric Car Stock.csv",
    "time_column": "Year",
    "value_columns": ["Total (millions)"],
    "dimension_columns": [],
    "frequency": "year"
  }
}

{
  "name": "time_series_arima",
  "arguments": {
    "series_collection": "<dataset from time_series_loader>",
    "order": [1, 1, 1],
    "forecast_steps": 3,
    "holdout_size": 3
  }
}

{
  "name": "time_series_exponential_smoothing",
  "arguments": {
    "series_collection": "<dataset from time_series_loader>",
    "trend": "add",
    "seasonal": null,
    "forecast_steps": 3,
    "holdout_size": 3,
    "random_seed": 42
  }
}

Possible result excerpt:

{
  "arima": {
    "series_id": "Total (millions)",
    "forecast_index": ["2026-12-31", "2027-12-31", "2028-12-31"],
    "forecast": [36.2602, 40.9644, 45.6134],
    "holdout_mape": 4.7453
  },
  "holt_winters": {
    "series_id": "Total (millions)",
    "forecast_index": ["2026-12-31", "2027-12-31", "2028-12-31"],
    "forecast": [36.3, 41.1, 45.9],
    "holdout_mape": 4.3264
  }
}

Example 3: Probabilistic zero-shot forecasting with Chronos-2

Prompt:

Load the Rossmann daily sales file, build daily sales series grouped by Store,
repair missing calendar gaps, and forecast the Sales series for the next
14 days with Chronos-2. Return 0.1, 0.5, and 0.9 quantiles, and show the
Store 1 result from the returned forecast list.

Tool calls:

{
  "name": "time_series_loader",
  "arguments": {
    "file_path": "docs/Rossmann Store Daily Sales.csv",
    "time_column": "Date",
    "value_columns": ["Sales"],
    "dimension_columns": ["Store"],
    "frequency": "day",
    "aggregation": "sum"
  }
}

{
  "name": "time_series_missing_data_handler",
  "arguments": {
    "series_collection": "<dataset from time_series_loader>",
    "strategy": "auto"
  }
}

{
  "name": "time_series_chronos2_forecast",
  "arguments": {
    "series_collection": "<dataset from time_series_missing_data_handler>",
    "target_metrics": ["Sales"],
    "forecast_steps": 14,
    "quantile_levels": [0.1, 0.5, 0.9],
    "context_length": 365
  }
}

Possible result shape:

{
  "ok": true,
  "forecast_results": [
    {
      "series_id": "Sales|Store=1",
      "model": "amazon/chronos-2",
      "forecast_index": ["2015-08-01", "2015-08-02", "..."],
      "forecast": ["<median forecast values>"],
      "quantiles": {
        "0.1": ["<lower quantile values>"],
        "0.5": ["<median quantile values>"],
        "0.9": ["<upper quantile values>"]
      },
      "inference_metadata": {
        "provider": "chronos-forecasting",
        "zero_shot": true,
        "prediction_length": 14,
        "context_length": 365
      }
    }
  ]
}

This example requires the chronos2 optional dependency and verified local Chronos-2 weights. target_metrics selects metric names, not dimensions; for dimension-specific inference, pass a prefiltered series_collection or filter the returned forecast_results by series_id.

Reproducibility And Operational Notes

• The server returns structured errors. Failures use the shape {"ok": false, "error": {"code": "...", "message": "...", "details": {}}}. Set DATA_SCIENCE_MCP_DEBUG=1 or TS_MCP_DEBUG=1 to include tracebacks.

• Statistical forecasting tools require complete target series. Run time_series_missing_data_handler before ARIMA, Holt-Winters, Chronos-2, or AutoML when missing values are present. Toto 2.0 can pass missing histories through target masks, but preprocessing is still useful for comparability.

• holdout_size computes validation metrics on a training fit, then refits on the full series for the returned future forecast.

• ARIMA supports known-future covariates through known_future_covariate_metrics, future_covariates, and calendar_covariates. AutoML and Chronos-2 also support past-only, known-future, and generated calendar covariates where the upstream model can use them.

• Canonical Chronos-2 and Toto 2.0 model weights are intentionally local and verified with resources/MODEL_WEIGHTS_MANIFEST.json. The canonical models do not silently fall back to remote downloads. For development only, set TS_MCP_ALLOW_REMOTE_MODELS=1 and explicitly request a non-vendored model ID.

• Runtime limits are configurable with environment variables such as TS_MCP_MAX_FILE_MB, TS_MCP_MAX_ROWS, TS_MCP_MAX_SERIES, TS_MCP_MAX_TIMESTAMPS, TS_MCP_CHRONOS2_MAX_CONTEXT, TS_MCP_TOTO2_MAX_CONTEXT, and TS_MCP_AUTOML_TIME_LIMIT_SECONDS.

• The implementation is organized in three layers: data_science_mcp/ for MCP server, CLI, registry, and errors; tools/ for MCP-facing adapters and rendering; and time_series_analysis/ for loading, diagnostics, quality handling, and forecasting logic.