← all repositories
Tejas-TA/predikit

Your sklearn model, now something an LLM can actually call

predikit wraps trained sklearn/XGBoost models as typed, schema-carrying tools for OpenAI, LangChain, or MCP — without the glue code.

predikit
Too new to rank. Velocity needs a few days of history.
not enough history yet
star history

What it does You hand it a fitted estimator plus a Pydantic input schema, and it produces a tool object with auto-generated JSON schema, validated inputs, and one-line exports: .to_openai(), .to_langchain(), .to_callable(). The LLM sees a named tool with a description; behind it, Pydantic v2 checks every argument before the model ever runs. There’s also a ToolRegistry for bulk export and an MCP server mode for serving a registry to MCP clients.

The interesting bit The boring details are the value: it maps inputs to features by name, not position, and raises a precise error telling you which schema fields don’t match the training columns. It also coerces LLM-isms like "yes"/"1" into real booleans before validation — the exact class of silent failure you’d otherwise debug at 2am.

Key highlights

  • ModelEnsemble runs multiple wrapped models in parallel with five reconciliation strategies (mean, vote, weighted variants, collect)
  • Confidence-aware routing: confidence_threshold plus warn/raise/fallback modes for classifiers with predict_proba
  • Loaders for MLflow and Snowflake model registries
  • Async support via ainvoke()
  • Optional CLI extras for predikit inspect and predikit serve

Caveats

  • Scikit-learn and XGBoost only — no PyTorch/transformers support
  • Field names in your Pydantic schema must exactly match training column names; numpy-trained models fall back to field order with no name checking
  • Confidence routing only works for classifiers exposing predict_proba

Verdict If you have tabular models sitting in a registry and agents that should be calling them, this removes a genuinely tedious layer. If your models are deep learning or your agents don’t need typed tool I/O, it’s not for you.

Frequently asked

What is Tejas-TA/predikit?
predikit wraps trained sklearn/XGBoost models as typed, schema-carrying tools for OpenAI, LangChain, or MCP — without the glue code.
Is predikit open source?
Yes — Tejas-TA/predikit is open source, released under the MIT license.
What language is predikit written in?
Tejas-TA/predikit is primarily written in Python.
How popular is predikit?
Tejas-TA/predikit has 856 stars on GitHub.
Where can I find predikit?
Tejas-TA/predikit is on GitHub at https://github.com/Tejas-TA/predikit.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.