OpenAI Agents SDK: setup and governed scope
Limited coverage · Runtime family · Reviewed 2026-09-07
Separately versioned Python and TypeScript managed tools with native approval callbacks and durable restart handling.
Set up OpenAI Agents SDK in your workspace Choose your first outcome
What is validated
Python managed HTTP tool
Limited coverage
Protected surfaces:
- operator-registered protected HTTP target
Authentication: user-backed personal credential, delegated user credential. Evidence capability: A; inspect each action's actual grade.
Setup instructions for this profile come from the SDK documentation your deployment operator provides.
TypeScript managed HTTP tool
Limited coverage
Protected surfaces:
- operator-registered protected HTTP target
Authentication: user-backed personal credential, delegated user credential. Evidence capability: A; inspect each action's actual grade.
Setup instructions for this profile come from the SDK documentation your deployment operator provides.
Start free, then install and connect
After the sample, use the local SDK installation guide. Obtain a matching source checkout or release artifact from your deployment operator; these pages do not assert that a package or marketplace entry has been published.
- Create a free workspace and verify your email. In your dashboard, choose Run free sample. Review the local rule’s allowed and blocked texts, saved audit ID, and reported signature availability. No model key or external target is needed. This sample does not connect an agent or create a target receipt or verified proof bundle.
- Return to your saved runtime setup and register an agent. Use separate runtime and personal review credentials; keep credentials outside source control.
- Choose an explicitly supported managed tool path. An administrator must configure a registered HTTP target or managed MCP connection and confirm your access.
- In a disposable workflow, exercise allow, deny, and revoked-authority cases. Record the actual action ID; a configured connection alone is not execution evidence.
- Inspect the event sequence, outcome, and evidence grade in the workspace. Use the evidence guide for SDK reads and independent verification.
Start free with the local rule sample, registration, and the controls available to your plan. Protected actions require proof.actions and the corresponding permission; exports require additional access. Ask your workspace administrator if these are unavailable. Provider usage may have its own charges.
Limits to test
- SDK-managed functions are covered by these profiles; provider-hosted tools and every Runner/handoff/session-store configuration need independent acceptance.
- Only the configured managed HTTP tool is covered; unrelated native tools, hosted tools and local memory are outside this profile.
- Packages are local release candidates; registry publication, production deployment and additional operating systems need separate acceptance.