Agent Registry — OpenAI Agents SDK
Wire a LumoAuth-registered agent into the OpenAI Agents SDK. The agent's identity is its registered AGENT_CLIENT_ID; it authenticates via OAuth 2.0 client credentials (via LumoAuthAgent); tools check has_capability(...) before running; MCP tools swap the base access token for an audience-bound MCP token using RFC 8693 — OAuth 2.0 Token Exchange.
Prerequisites
Install the lumoauth package and set the environment variables
LUMOAUTH_URL, LUMOAUTH_ORG_ID, AGENT_CLIENT_ID, and
AGENT_CLIENT_SECRET. See Install the SDK.
Install
pip install lumoauth openai-agents
Example
import asyncio
from agents import Agent, Runner, function_tool
from lumoauth import LumoAuthAgent
# 1. Initialize and authenticate the LumoAuth Agent
lumo_agent = LumoAuthAgent()
lumo_agent.authenticate()
# 2. Define tools that check LumoAuth capabilities
@function_tool
def search_company_documents(query: str) -> str:
"""Search internal company documents for the given query."""
if not lumo_agent.has_capability('read:documents'):
return "Error: Agent lacks 'read:documents' capability."
return f"Found 3 documents matching '{query}'"
@function_tool
def query_financial_mcp(metric: str) -> str:
"""Query the secured financial metrics MCP server for a given metric."""
if not lumo_agent.has_capability('mcp:financial'):
return "Error: Agent lacks 'mcp:financial' capability."
mcp_token = lumo_agent.get_mcp_token("urn:mcp:financial-data")
if not mcp_token:
return "Error: Failed to obtain token for Financial MCP server."
return f"Retrieved {metric}: $1.2M (Authenticated via MCP Token Exchange)"
# 3. Check budget before running
if lumo_agent.is_budget_exhausted():
raise RuntimeError("Agent budget exhausted for today.")
# 4. Create the OpenAI Agent
research_agent = Agent(
name="Research Analyst",
instructions="You are a senior analyst. Use tools to search documents and query financial metrics.",
tools=[search_company_documents, query_financial_mcp],
)
# 5. Run the agent
async def main():
result = await Runner.run(
research_agent,
"Search documents for Q3 performance, then query the financial MCP for EBITDA.",
)
print(result.final_output)
if __name__ == "__main__":
asyncio.run(main())
How it works
| Step | What happens |
|---|---|
LumoAuthAgent() | Reads credentials from env vars and initializes the client |
lumo_agent.authenticate() | Runs the OAuth 2.0 client-credentials flow |
@function_tool | Decorates plain Python functions as OpenAI SDK tools |
has_capability(...) | Guards tool execution with LumoAuth capability checks |
get_mcp_token(...) | Swaps the agent token for an MCP-specific token (RFC 8693) |
Next steps
- Agent Registry overview — manage registrations, budgets, and capabilities
- JIT Permissions — OpenAI Agents SDK — add just-in-time permission escalation