predikit
A library that wraps a trained scikit-learn or XGBoost model as an LLM-callable tool, generating a JSON schema and validated types automatically so agent frameworks can call it directly.
predikit wraps a fitted scikit-learn or XGBoost model so an LLM agent can call it as a typed, schema-validated tool. It auto-generates a JSON schema from a data model describing the tool's inputs, validates and casts values, and exposes one-line conversions to OpenAI function-calling and LangChain tool formats, plus an async invoke method.
It also includes model ensembling with several built-in aggregation strategies, confidence-threshold handling with configurable behavior on low confidence, and loaders that pull models directly from an MLflow or Snowflake model registry instead of requiring manual registry calls. A command-line interface and an MCP server are also provided.
It is written in Python and MIT licensed.