MCPStack AutoDDG MCP

Automatic dataset topics & descriptions — powered by AutoDDG and MCPStack

Video Trailer Showing The MCP In Action: https://www.youtube.com/watch?v=Wer8y8FWwM8

Important

If you haven’t visited the MCPStack main orchestrator repository yet, please start there: MCPStack

Caution

Please be aware that this MCP is in an early-alpha stage. While it is functional and can be used for various tasks, it may still contain bugs or incomplete features. Feel free to report any issues you encounter or suggest improvements. Even better, feel free to contribute directly!

Warning

Please be aware that you cannot use this MCP without an OpenAI-compatible API key. To gen. one, please visit: https://platform.openai.com/account/api-keys

Note

For the time being, this MCP is working with the branch feat/modern_pythonic_library_transformation from the mother library, AutoDDG. See more at: VIDA-NYU/AutoDDG#4. As such, we recommend you to install AutoDDG from source with this library, until the PR is merged upstream. Adapt the "autoddg" in the dependencies in pyproject.toml accordingly.

💡 About The MCPStack AutoDDG Tool

This repository provides a native MCP around the AutoDDG library for dataset description and discovery:

• Load a CSV and keep a deterministic sample (by size or percentage).
• Profile a dataframe (datamart-style notes).
• Infer a semantic profile for columns.
• Generate a concise topic.
• Produce a readable dataset description.
• Expand that description for search/discovery (tune the temperature etc.).
• Optionally evaluate the description with a separate evaluator key.

AutoDDG official library (without the MCP wrapper): https://github.com/VIDA-NYU/AutoDDG

Installation

The tool is distributed as a standard Python package. MCPStack will auto-discover it.

Via uv (recommended)

uv add mcpstack-autoddg

Via pip

pip install mcpstack-autoddg

(Dev) Pre-commit hooks

uv run pre-commit install
# or: pre-commit install

Using With MCPStack — CLI workflow

This tool declares entry points so MCPStack can see it automatically:

[project.entry-points."mcpstack.tools"]
autoddgtool = "mcpstack_autoddg.tool:AutoDDGTool"

[project.entry-points."mcpstack.tool_clis"]
autoddgtool = "mcpstack_autoddg.cli:AutoDDGCLI.get_app"

1) (Optional) Configure environment

AutoDDG requires an OpenAI-compatible key. You may optionally provide a separate evaluator key:

AUTO_DDG_OPENAI_API_KEY: "<your key>" (required for generation)
AUTO_DDG_EVALUATOR_API_KEY: "<your key>" (optional; falls back to AUTO_DDG_OPENAI_API_KEY)

Use the CLI to generate a config file (useful for CI or sharing defaults):

uv run mcpstack tools autoddg configure
# Then is followed an interactive prompt to config and set parameters.

Or you can pass parameters directly, e.g.:

uv run mcpstack tools autoddg configure \
  --model-name gpt-4o-mini \
  --description-words 120 \
  --description-temperature 0.0 \
  --topic-temperature 0.0 \
  --api-key sk-... \
  --evaluator-key sk-... \
  -o autoddg_config.json \
  --verbose

For others, feel free to uv run mcpstack tools autoddg --help to see all options.

2) Add to a pipeline

Create or extend a pipeline with AutoDDG:

# New pipeline
uv run mcpstack pipeline autoddg --new-pipeline my_pipeline.json --tool-config autoddg_config.json

# Or append to an existing one
uv run mcpstack pipeline autoddg --to-pipeline my_pipeline.json --tool-config autoddg_config.json

Programmatic API Workflow

Use the AutoDDG tool directly in a stack:

from MCPStack.stack import MCPStackCore
from mcpstack_autoddg import AutoDDGTool

pipeline = (
    MCPStackCore()
    .with_tool(AutoDDGTool(
        model_name="gpt-4o-mini",
        search_model_name=None,
        semantic_model_name=None,
        description_words=120,
        description_temperature=0.0,
        topic_temperature=0.0,
        evaluator_model_name="gpt-4o",
    ))
    .build(type="fastmcp", save_path="autoddg_pipeline.json")
    .run()
)

AutoDDG Actions Supported

Note

If any action fails, feel free to open an issue so we may update with the potential changes on the mother library, AutoDDG. https://github.com/VIDA-NYU/AutoDDG

• load_dataset(csv_path|csv_text, sample_size?, sample_percent?, random_state=9) → load CSV and store a sampled CSV string in state
• profile_dataset() → datamart-like profile; may also return semantic notes
• generate_semantic_profile() → infer semantic metadata for columns
• generate_topic(title, original_description?, dataset_sample?) → concise dataset topic
• generate_description(dataset_sample?, use_profile=True, use_semantic_profile=True, use_topic=True) → readable description; enforces prerequisites if the flags are left on
• expand_description_for_search() → search-oriented variant of the last description (needs a topic)
• evaluate_description() → runs evaluator (requires evaluator key or reuse of generation key)
• get_state_summary() → booleans for which artifacts exist in state

License

MIT — see LICENSE.