Agent Adapter: plain Python functions as ADK tools, with one-use approvals and budgets

[Veloce-AI]

Hi all, I'm one of the authors of Agent Adapter, an Apache-2.0 Python library (pip install "agent-adapter[adk]").

What it does for ADK users:

• adk_tools() turns plain Python functions into ADK BaseTools. The same functions also run in LangChain, CrewAI, OpenAI Agents SDK and 17 more, or over MCP.
• A confirm=True tool pauses with ADK's adk_request_confirmation and runs only after your approver issues a one-use token bound to that call and its arguments.
• Every call goes through one executor with budgets, a cost ledger, guards, caching and OpenTelemetry spans.

registry = Registry()

@registry.tool(confirm=True)
def send_invoice(customer_id: str, amount: float) -> str:
    return f"invoiced {customer_id} for {amount:.2f}"

tools = adk_tools(registry, executor=Executor(registry, approval=port))
agent = LlmAgent(name="billing", model="gemini-2.5-flash", tools=tools)

port is your approver: a callback, or a SQLite-backed DeferredApproval that survives a restart. A Flow graph can also compile to an ADK Workflow.

Docs and the full runnable example: https://agent-adapter.veloceai.in/integrations/adk/

How do you handle human approval for ADK tool calls today? Would a token bound to the exact call fit your setup?