"fake_search is not a valid tool, try one of [tool]" in LangGraph

"Error: my_tool is not a valid tool, try one of [tool]" in LangGraph

TL;DR: your tools are registered in exactly one of the two places they need to be. Pass the tool list to BOTH llm.bind_tools(tools) AND ToolNode(tools). The model can only call tools it knows about, and the ToolNode can only execute tools it was given. Miss either one and you get this error.

Assistant: Error: fake_search is not a valid tool, try one of [tool]

When this applies

• Built-in tools (Tavily, DuckDuckGo) work, but YOUR custom function gets rejected.
• The error names your function and lists the tools that ARE valid.
• You followed a tutorial and swapped in your own tool.

When it does not

• If NO tools work, check your imports and that the tools list is not empty.
• ValidationError when DEFINING the tool means the function signature is bad. Different fix.

Fix it

1. Register the tool in both places

from langgraph.prebuilt import ToolNode

tools = [tavily_tool, duckduckgo_tool, fake_search]  # your custom tool included

# place 1: the model must know the tool exists
llm_with_tools = llm.bind_tools(tools)

# place 2: the ToolNode must be able to execute it
tool_node = ToolNode(tools)

Expected: the model emits a fake_search tool call, and the ToolNode executes it instead of erroring.

2. Check the name matches exactly

from langchain_core.tools import tool

@tool
def fake_search(query: str) -> str:
    """Search a fake index."""
    return "result"

print(fake_search.name)  # 'fake_search' -- this is the name the model must emit

Expected: the name the model calls matches tool.name. A mismatch (e.g. the model says fakeSearch) means the bind_tools schema did not reach the model.

3. If you defined the tool in another file, verify the import

# the classic tutorial-swap bug: you imported the tutorial's tools list
# instead of your own
from my_tools import fake_search   # make sure THIS is what lands in `tools`

Expected: tools actually contains your function, not just the tutorial's examples.

Why it happens

LangGraph splits tool use in two: bind_tools teaches the MODEL the tool schemas (so it emits the right tool calls), and ToolNode(tools) gives the EXECUTOR the implementations. Tutorials often define tools = [tavily] once and pass it to both; when you add your custom function you update one call site and forget the other. The model then either never calls your tool (bind_tools missed) or calls it and the executor rejects it (ToolNode missed), which is this error.

Edge cases

• create_react_agent(model, tools=[...]) does both registrations for you. If you keep hitting this with a manual graph, consider switching.
• Tool names must be valid identifiers for most models: no spaces, no special characters.
• After changing the tools list, restart the notebook kernel. Stale bind_tools closures are a real gotcha.

Compatibility

All langgraph versions with ToolNode and bind_tools (0.x and 1.x, Python).