OpenAI Agents SDK - Danube

Install

pip install "danube[openai-agents]"
export DANUBE_API_KEY=...   # from the dashboard, or POST /v1/agents for a self-registered agent

Use

from agents import Agent, Runner
from danube import DanubeClient
from danube.integrations.openai_agents import danube_tools

client = DanubeClient()  # reads DANUBE_API_KEY

agent = Agent(
    name="assistant",
    instructions="Use Danube to find and call tools. Search before assuming a capability exists.",
    tools=danube_tools(client),
)

result = Runner.run_sync(agent, "What are the top three Hacker News stories right now?")
print(result.final_output)

danube_tools(client) returns three FunctionTools:

tool	what the agent gets
`danube_search_tools(query, limit)`	tool ids, descriptions and parameter schemas from the whole catalog, with `readiness` and `reliability` when the backend stamps them
`danube_describe_tool(tool_id)`	the full schema, tips from earlier callers, readiness for this account
`danube_execute_tool(tool_id, parameters)`	the tool’s result, or an error that says what to do (connect a service, fix an argument)

Options: service_id to limit search to one service, ready_only=True to hide tools this account has not connected, max_results for the search page size. The Danube client is synchronous; the adapter runs it in a worker thread so the agent loop stays responsive.

Why three tools rather than one per catalog entry

The catalog has thousands of tools; putting them all in the model’s context is slow and expensive, and the set changes daily. Search first keeps the prompt small and always current, and describe gives the model the schema only for the tool it is about to call.