Skip to main content

Building AI Agents with IterableData

Canonical guide for integrating IterableData with LLM agents, IDEs, and automation.

Quick start​

from iterable.tools import detect_format, read_sample, infer_schema

# Uniform JSON envelopes
fmt = detect_format("data.csv")
rows = read_sample("data.csv", n=10, redact=True)
sch = infer_schema("data.csv")

Install extras as needed:

pip install iterabledata[ai]        # documentation generation
pip install iterabledata[langchain] # LangChain StructuredTools
pip install iterabledata[mcp]       # MCP server (iterable-mcp)

Layers​

  1. iterable.tools — stable tool functions with ok / data / error envelopes
  2. iterable.tools.schemas — OpenAI, Anthropic, and JSON Schema exports
  3. iterable.catalog — format metadata for routing (describe_format("xml"))
  4. iterable.ai.context — sample_for_llm, redact_for_llm before cloud APIs
  5. iterable-mcp — stdio MCP server for Cursor / Claude Desktop

OpenAI function calling​

from iterable.tools import schemas

functions = schemas.to_openai_functions()
# Pass to client.chat.completions.create(..., tools=functions)
result = schemas.call_tool("read_sample", {"path": "data.csv", "n": 5})

See examples/ai/openai_function_calling.py.

Anthropic tools​

tools = schemas.to_anthropic_tools()
result = schemas.call_tool("infer_schema", {"path": "data.csv"})

See examples/ai/anthropic_tools.py.

LangChain​

from iterable.tools.langchain import get_tools

MCP (Cursor, Claude Desktop)​

Configure MCP to run iterable-mcp after pip install iterabledata[mcp]. See MCP.md.

Safety​

  • Use read_sample + redact_for_llm before sending data to cloud LLMs
  • convert_file requires confirm=True to write files
  • Do not exec() LLM-generated code; use explicit pipeline() transforms

Further reading​