Choose one tool call that can become paid usage or a vendor action
Do not start by wrapping the whole runtime. Start with the one call that can create cost or consequence today.
NORNR
Govern one consequential tool call before it becomes normal runtime behavior.
NORNR / OpenAI runtime gate
Productized packagePut one OpenAI Agents tool call behind policy before it can create spend.
This package is the cleanest runtime story for teams already using OpenAI Agents SDK. Pick one consequential tool call, make NORNR decide whether it should run, and keep one defended record when it does.
The right first runtime package is not “govern every tool.” It is one paid or consequential tool lane with one decision path, one review path and one defended record.
Package shape
One runtime spend gate, not a framework-wide rewrite.
The package should be easy to understand, easy to start and easy to prove in front of a buyer.
Approved, queued or blocked before the runtime clears the action
The model can still reason. NORNR only decides whether this tool action may continue under the current owner mandate.
One operator or finance reviewer owns the queue above threshold
The package becomes buyer-safe when the escalation owner is explicit and the reason lands in one queue item.
The tool call should still end in one exported record later
Runtime-native only matters if finance, risk and audit can still reuse the same path afterward.
Install-first path
Make the OpenAI lane feel official from the first command.
Starter, pack, guide chain and outputs should all line up without the team having to improvise the posture model.
Copy-paste the official OpenAI lane
bash
nornr init openai-agents --owner runtime-agentopenai-tool-governed
Start with one threshold, one counterparty posture and one explicit review owner for the single consequential tool lane.
Install first, then narrow the runtime story
One governed tool lane should already prove the full stack
What it proves
OpenAI Agents can stay native while NORNR becomes the control layer above them.
That is the commercial point of the package: keep the runtime, change the decision and proof posture.
Agent proposes one tool action
Amount, purpose and counterparty enter before the call runs.
NORNR evaluates the mandate
The tool call is checked against threshold, counterparty and policy posture.
Runtime clears or queues
Low-risk cases stay fast. Higher-risk cases route into review without breaking the runtime shape.
Review resolves from one queue
The reviewer sees the reason, not a separate reconstruction.
Execution and receipt trail stay attached
If approved, the resulting action still lands in the same governed record.
Finance exports the same path later
The package is complete only when close and audit still read cleanly afterward.