IronClaw: Agentic API¶

IronClaw is the agentic interface for GlassBox, designed specifically for WASM-based environments (like Pyodide). It provides a high-level, stateless API that allows LLM agents to orchestrate complex AutoML pipelines.

Key Features¶

Stateless Orchestration: Enables independent tool calls by managing an in-memory state.
Agent-Ready Schemas: Functions are designed to be easily invoked by LLMs with clear input/output structures.
WASM Optimized: Lightweight and compatible with browser-based Python execution.

Architecture¶

The following diagram illustrates how IronClaw interacts with the core GlassBox library and the Agent:

Kroki

Main API Functions¶

`inspect_data_api(csv_filename)`¶

Loads a CSV file into memory and generates a comprehensive EDA report.

`clean_data_api(...)`¶

Applies a suite of cleaning operations (imputation, encoding, scaling, etc.) based on agent-defined strategies.

`train_and_tune_api(...)`¶

Orchestrates model selection and hyperparameter optimization to find the best performing model for a target column.

State Management¶

Since web-based agents often call tools in a loop where Python state might be reset or handled carefully, IronClaw uses a global StateManager to retain pointers to the active dataset.

from glassbox.ironclaw.state import memory

# Inspect first to load into memory
inspect_data_api("data.csv")

# Subsequent calls automatically find the dataset
clean_data_api(...)