Tool Calling

@tool

Slap @tool on a function. LangChain picks up the name from the function, the description from the docstring, and the argument types from your hints. Your code still runs as plain Python — return a string the model can use on the next pass.

Keep the function name in snake_case (that is what OpenAI sees). Type every parameter. One clear docstring line beats a paragraph.

from langchain.tools import tool

TAGS = {
    "a": "Creates a hyperlink. Use href for the URL.",
    "title": "Sets the browser tab title.",
}

@tool
def lookup_html_tag(tag_name: str) -> str:
    """Look up a short description of an HTML tag by name."""
    key = tag_name.strip().lower().lstrip("<").rstrip(">")
    return TAGS.get(key, f"No entry for '{tag_name}'.")

print(lookup_html_tag.name)        # lookup_html_tag
print(lookup_html_tag.description) # first line of docstring
print(lookup_html_tag.args)        # schema sent to the model

What OpenAI gets

Passing tools=[lookup_html_tag] sends JSON like below — not the function body. The demo prints lookup_html_tag.args first so you can sanity-check before calling the API.

name: lookup_html_tag
description: Look up a short description of an HTML tag by name.
parameters:
  tag_name: string  (required)

tool_calls on one question

Sometimes the assistant reply is only a tool_calls block, no user text yet. LangChain calls lookup_html_tag, pushes the return value back in, and asks the model again for the final sentence.

Same create_agent call as multiply — swap the tool:

from langchain.agents import create_agent

agent = create_agent(model="openai:gpt-4o-mini", tools=[lookup_html_tag])

result = agent.invoke({
    "messages": [{"role": "user", "content": "What does the <a> tag do? Use lookup_html_tag."}],
})
print(result["messages"][-1].content)

Run the demo

Venv from Project Setup. Download, then:

python tool_calling_demo.py

Custom tool names and Pydantic args later: Tools docs.

What's Next

Next: ReAct Framework.