Namespace: openaiAgents
Classes
- OpenAIAgentsPlugin
- OpenAIAgentsTraceClientInterceptor
- SandboxClientProvider
- StatefulMCPServerProvider
- StatelessMCPServerProvider
Interfaces
- HostedToolCredentials
- HostedToolDomainSecret
- ModelActivityOptions
- ModelSummaryProvider
- OpenAIAgentsPluginInterceptorOptions
- OpenAIAgentsPluginOptions
- OpenAIAgentsTraceClientInterceptorOptions
Type Aliases
HostedToolCredentialsResolver
Ƭ HostedToolCredentialsResolver: (identity: HostedToolIdentity) => HostedToolCredentials | undefined | Promise<HostedToolCredentials | undefined>
Resolves the credentials for one hosted tool, Worker-side.
Called for every hosted tool that has somewhere to put a credential, on each
model invocation, so a deployment that reads from a secret manager should
cache. A hosted MCP tool needs a server label to be asked about; a shell tool
needs a container_auto environment with a domain allowlist; a code
interpreter needs an inline container definition with a domain allowlist —
naming an existing container by id skips the callback even though that
container may carry an allowlist of its own. Only the fields the Workflow left
out are filled in; return nothing to leave a tool as the Workflow declared it.
A throw fails the model Activity, keeping whatever retryability it carries. The error's own message reaches Workflow history, so keep credentials out of it.
Type declaration
▸ (identity): HostedToolCredentials | undefined | Promise<HostedToolCredentials | undefined>
Parameters
| Name | Type |
|---|---|
identity | HostedToolIdentity |
Returns
HostedToolCredentials | undefined | Promise<HostedToolCredentials | undefined>
HostedToolIdentity
Ƭ HostedToolIdentity: { connectorId?: string ; name: string ; serverLabel: string ; serverUrl?: string ; tool: "hostedMcp" } | { allowedDomains: string[] ; name: string ; tool: "shell" } | { allowedDomains: string[] ; name: string ; tool: "codeInterpreter" }
The hosted tool a credential is being requested for, identified by the
non-secret configuration the Workflow declared. A shell or code interpreter
tool carries its own name; every tool hostedMcpTool() builds is named
hosted_mcp, so serverLabel with serverUrl or connectorId is what
separates two of those — two sharing both are indistinguishable here.
MCPServerProvider
Ƭ MCPServerProvider: StatelessMCPServerProvider | StatefulMCPServerProvider
SerializableModelActivityOptions
Ƭ SerializableModelActivityOptions: Omit<ModelActivityOptions, "summary"> & { summary?: string }
Plugin-side model Activity options propagated to Workflows via the
__openai_agents_config header. The function form of summary
is excluded because functions cannot survive JSON serialization.
Variables
DEDICATED_WORKER_FAILURE_TYPE
• Const DEDICATED_WORKER_FAILURE_TYPE: "DedicatedWorkerFailure"
Error type for dedicated-Worker scheduling and heartbeat failures.