How do I choose the OpenAI Agents SDK Montage execution boundary?
Before adapting connection code, decide whether the application or Responses API should execute remote tools. A locally controlled Montage session cannot be assumed to transfer its authorization, lifecycle or approval rules into a hosted tool configuration, even when both target the same endpoint. Compare the two execution arrangements before choosing parameters. MCPServerStreamableHttp owns a local connection lifecycle, whereas HostedMCPTool delegates execution through Responses; shared endpoints do not imply shared consent, approval settings or verification evidence.
Help me choose the OpenAI Agents SDK Montage execution boundary. Identify whether the existing application opens MCPServerStreamableHttp or declares HostedMCPTool for Responses execution. Record local status evidence as local only. No writes.Map the current caller
Identify whether the existing application opens MCPServerStreamableHttp or declares HostedMCPTool for Responses execution. Do not treat these as interchangeable wrappers around one local session. The boundary inventory should distinguish a locally owned session from a HostedMCPTool declaration, even if both configurations contain the same remote endpoint.
Name the required data boundary
Determine where tool inputs and outputs are intended to travel. Keep a local diagnostic distinct from a decision to send account-scoped workspace data through a hosted execution path. An execution-boundary review can identify data movement without making a paid Responses request or changing the application’s current deployment architecture.
Separate authorization responsibilities
Review approved Montage auth for the chosen execution arrangement. An OpenAI API credential authorizes OpenAI services and cannot stand in for a Montage account grant. Local OAuth state must not be assumed to appear automatically in hosted execution; each path needs its own supported approved authorization configuration.
Use a direct preliminary read
Validate the locally controlled connection with signals_status without starting an agent or Responses call. This establishes a narrow local baseline without incurring inference merely to diagnose transport. The direct baseline should use the intended account and return actual workspace status, keeping a model-generated summary out of the acceptance criterion.
Preserve the unverified hosted state
Record local status evidence as local only. Any hosted execution needs its own explicitly approved configuration and verification rather than inheriting a success claim from the direct connection. A hosted follow-up remains unverified until explicitly approved and exercised; the local receipt must not be re-labeled as proof of hosted authorization or approval parity.
Keep the boundaries clear.
- Local OAuth state must not be assumed to appear automatically in hosted execution; each path needs its own supported approved authorization configuration.
- A hosted follow-up remains unverified until explicitly approved and exercised; the local receipt must not be re-labeled as proof of hosted authorization or approval parity.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.