Integrum

Convert any Python module or library into a FastMCP Model Context Protocol server.

Integrum turns any Python module, library, function, or class into a Model Context Protocol (MCP) server powered by FastMCP.

Integrum uses Python reflection (inspect), regex filtering, and an in-memory Word-Based Object Store (playful_llama, swift_falcon) so that object-oriented libraries (such as scikit-learn, pandas, PyTorch) can be used naturally by LLMs across multi-turn conversations without token bloat.

Features

• Module Reflection: Automatically introspects functions and classes using inspect.
• Regex Filtering: Expose only what you need with --include and --exclude regex patterns.
• Word-Based State Management: Object IDs are short, memorable words (e.g. playful_llama) to save tokens and avoid UUID hallucination.
• Multi-User Isolation: Supports FastMCP Context sessions to partition objects per connecting client.
• Docstring Instruction Injection: Transforms Python docstrings into LLM tool documentation and automatically injects state-handling instructions.
• Dot-Notation Attribute Resolution: Pass swift_falcon.data or swift_falcon.target directly to subsequent tool calls like fit().
• CLI & Python API: Run from the command line (integrum expose) or integrate directly into Python scripts.

Installation

uv pip install integrum
# or
pip install integrum

Quick Start (CLI)

Expose scikit-learn dataset loaders and RandomForestClassifier as an MCP server:

integrum expose sklearn.datasets.load_iris sklearn.ensemble.RandomForestClassifier

Preview what tools will be exposed:

integrum list sklearn.datasets.load_iris sklearn.ensemble.RandomForestClassifier

Filter functions and methods with regex:

integrum expose sklearn.datasets sklearn.ensemble \
  --include "^(load_iris|RandomForestClassifier)$"

Run with HTTP or SSE transport:

integrum expose sklearn.datasets.load_iris --transport http --port 8000

Python API

from integrum import IntegrumServer

server = IntegrumServer(name="ml-server")
server.expose(["sklearn.datasets.load_iris", "sklearn.ensemble.RandomForestClassifier"])

# Run over stdio (default for MCP desktop clients)
server.run(transport="stdio")

How State Management Works

When a class is exposed, Integrum registers:

1. <ClassName>_create: Instantiates the class and stores it under a word ID (e.g., playful_llama).
2. <ClassName>_<method>: Accepts a virtual parameter obj_id: str (e.g., obj_id="playful_llama"), resolves the underlying instance, and executes the method.

Returned objects (such as datasets or numpy arrays) are also registered in the store with word IDs, allowing subsequent tools to receive them as arguments.

Examples

Check out examples/gemma_sklearn_session.md and examples/gemma_sklearn_session.json for a complete, verified multi-turn transcript of Gemma 4 26B loading the Iris dataset, performing train_test_split, training a RandomForestClassifier, and evaluating accuracy.

Links & Homepage

• Homepage & Repository: https://github.com/alphadeepai/integrum
• Issue Tracker: https://github.com/alphadeepai/integrum/issues

License

This project is licensed under the terms of the MIT License. Copyright (c) 2026 AlphaDeep.