AutoML MCP

What is AutoML MCP?

AutoML MCP is a specialized toolkit designed to empower AI assistants (Agents) to execute end-to-end Machine Learning projects in a deterministic and structured way.

By providing a Model Context Protocol (MCP) server alongside a dedicated CLI, it gives AI agents the exact utilities they need to handle the entire ML lifecycle—from initial exploratory data analysis to model training and deployment.

What it IS:

• A bridge that gives AI agents structured file management, Jupyter notebook manipulation, and dashboard creation capabilities.
• A standardized workflow enforcer for ML projects created by AI.
• A toolset that allows agents to safely register datasets, manage pipeline stages, and visualize data.

What it IS NOT:

• It is not a version control system like Git.
• It is not a general project or package manager like uv or poetry (It manages the files and items created by agent or given by the user required in the Machine Learning Workflow).

---

Installation & Usage

Prerequisites

Make sure you have uv installed.

Installation

Using uv tools:

uv tool install git+https://github.com/legit-os/AutoML_MCP_Server.git -U

Basic Workflow

1. Initialize a new ML project directory:

   automl init ./my_ml_project -n MyProject

2. Start the MCP Server (so your AI agent can connect):

   automl mcp start

   (To get the config JSON to paste into your MCP client, run automl mcp)
3. Launch the Dashboard (to view agent-generated visualizations):

   automl dashboard

---

What We Offer

1. The automl CLI

The CLI is used by the human user to manage projects and track state. Key commands include:

Command	Description
automl init <dir>	Initializes a directory for tracking and creates the base folder structure.
automl set <name>	Sets the active working project for the MCP server.
automl get	Shows the currently active project.
automl dashboard	Launches the React/FastAPI dashboard to view visualizations.
automl mcp	Prints the JSON configuration for your MCP client.
automl mcp start	Starts the FastMCP server over HTTP.
automl track <type>	Manually syncs or registers existing files into the project configuration.

2. The MCP Server Tools

The MCP Server exposes these tools directly to the AI agent, allowing it to manipulate the project dynamically. More than 20 tools are exposed in this mcp server, ensure your agent has enough context and your llm is trained to pick correct tool from too many tool options. If you use a agent that uses an extra tool calling agent or dedicated tool like the agent of antigravity, then it is OK.

Tool Name	Description
get_current_project_info	Reads the project configuration, showing active datasets, files, and paths.
manage_plan	Reads or writes to agentplan.md to establish a plan before coding.
write_pipeline_element	Creates modular .py files for specific ML pipeline stages (e.g., preprocessing, training).
write_util	Creates utility scripts for the pipeline.
manage_notebook	Reads, adds, edits, or deletes specific cells in Jupyter Notebooks (.ipynb).
list_notebooks	Lists all notebooks currently in the project.
register_dataset / delete_dataset	Tracks or removes dataset references in the project config.
manage_ops_file	Writes infrastructure files (Dockerfiles, DVC configs, etc.).
create_analysis_dashboard_item	Writes an analysis script and automatically captures selected variables (DataFrames, plots) to the dashboard.
read_dashboard_items	Allows the agent to read back the data (lists, dicts, DataFrames) it previously plotted on the dashboard.
update_project_metadata	Updates global project metadata (author, version, etc.).

This server doesn't have a tool to run cli commands since many agents restrict that behaviour, you may need to use your agent providers native mcp server for running commands (like antigravity and claude can run commands from there own capabilties).