Custom Component as Tool

Where we left off

Lesson 12's TitleCaseComponent had exactly one place it could go, wired directly into the next component's input, always run, every time. A tool is different: it's offered to an Agent, which decides for itself, per request, whether to call it at all. This lesson writes the same kind of Custom Component again, but shaped to be a tool instead of a fixed step in the chain.

Why this lesson runs the graph directly

Same reason as Lesson 12: graph.arun() runs the in-memory graph without any JSON deserialization step, so the migration-loop bug from Lesson 2 (triggered only by loading an unrecognized component type from JSON) never enters the picture. flow.json here is for reference only, not for reloading through run_flow_from_json or the REST API.

Do this yourself

1. Build on Lesson 4's flow (Google Generative AI already on the canvas) or start fresh: Chat Input, Google Generative AI, and an Agent component (search "Agent" in the sidebar).
2. Connect Google Generative AI's Model output (not its Message output, a different port, this one has type LanguageModel) into the Agent's Model input. This is a different wiring than every earlier lesson, the Agent consumes a model as a connection, not a typed-in name.
3. Add a New Custom Component the same way as Lesson 12, paste in WordLengthTool from lesson.py below. Connect its Tool output into the Agent's Tools input.
4. Connect Chat Input into the Agent's Input field, and the Agent's Response output into a Chat Output.
5. In the Playground, ask something like "how many letters are in 'xylophone'?" and something unrelated like "what's the capital of France?". Watch the Agent's own reasoning trace in the Playground, it should call the tool for the first question and answer directly for the second.

No canvas.png ships with this lesson: re-importing a saved flow that wires a model into an Agent's Model field this way, rather than connecting it by hand on the canvas, trips a frontend validation quirk in this Langflow version, the import silently drops that one edge and shows "Some connections were removed because they were invalid." The flow still runs correctly (verified via lesson.py and the REST API, this is purely a re-import display quirk), but a screenshot produced that way would show the Agent's Model input disconnected, which would be actively misleading here. Build it by hand following the steps above and it wires and runs exactly as described.

The code, piece by piece

def build_tool(self) -> Tool:
    from langchain_core.tools import StructuredTool
    ...
    return StructuredTool.from_function(
        name="word_length",
        description="Return the number of characters in a single word.",
        func=_word_length,
        args_schema=WordLengthSchema,
    )

Lesson 12's output method returned a Message, meant to feed straight into the next component. This one returns a LangChain Tool object, meant to be handed to an Agent, which reads its name and description to decide when it's relevant, the same object you'd build if you were writing this tool for a raw LangChain agent with no Langflow involved at all.

agent.set(
    model=gemini.build_model,
    tools=[word_length.build_tool],
    ...
)

model=gemini.build_model connects the Google Generative AI component's model output, not its text output, into the Agent, this is what lets the Agent do its own reasoning and tool-calling with that model. tools=[word_length.build_tool] is a list, an Agent can be given several tools this way, each entry another component's tool output method.