mloda

# mloda: Declarative Data Access for AI Agents ## Check the registry index first Always do this before hand-writing a chain step. mloda-registry's community set covers common `{col}__{op}` transforms: aggregation, window/scalar/frame aggregate, scalar/point arithmetic, rank, offset, percentile, binning, datetime, string ops, time bucketization, ffill, ema, sessionization, resample. None of it ships with plain `pip install mloda`; install what you need, e.g. `pip install mloda-community` (all) or `pip install "mloda-community-rank[pandas]"` (one). See its [plugins table](https://github.com/mloda-ai/mloda-registry#plugins) for the full list and feature-name patterns. The table drifts, so once installed, confirm what's actually loaded rather than trusting a static list: ```python from mloda.user import PluginLoader from mloda.steward import get_feature_group_docs PluginLoader.all() for fg in get_feature_group_docs(): print(fg.name, fg.description) ``` This only reflects installed packages: an empty or unrelated result means the plugin isn't installed yet, not that it doesn't exist. Nothing covering it? See the `mloda-plugins` skill for how to write a FeatureGroup, ComputeFramework, or Extender. ## Mental model mloda separates **WHAT** a caller needs from **HOW** it gets computed. A caller (human or LLM) requests features by name; mloda resolves dependencies across a plugin graph and executes them on a compute framework (`PandasDataFrame`, `PyArrowTable`, `PythonDict...