Should my application or the Responses API own the MCP connection?
The OpenAI Agents SDK documents two different remote MCP execution arrangements: MCPServerStreamableHttp for application-controlled connections and HostedMCPTool when the Responses API executes remote tools. Choose that boundary before configuring credentials or tool approval. This guide does not run paid inference, and an OpenAI API credential must never be confused with authorization to a Montage account or workspace.
https://app.montages.co/mcpUsing the authorized local MCP connection, execute signals_status directly and inspect the real workspace result. Do not initiate an agent run or a hosted Responses call for this check.Select the execution boundary
Use MCPServerStreamableHttp when your application should manage the remote connection and its lifecycle locally. Consider HostedMCPTool only when you intentionally want Responses API execution of the remote tool. These are not interchangeable wrappers around the same local session. Review the existing application's SDK version and data boundaries before selecting either design.
Set the Streamable HTTP address
Configure the selected documented interface with https://app.montages.co/mcp using the parameter schema for the installed SDK. A locally controlled MCP server object still reaches this hosted endpoint; it does not require running a Montage server process. Do not copy a stdio subprocess example or assume a hosted tool automatically inherits your local client's session state.
Separate the two authorization systems
The application owns approved Montage authorization for its chosen execution path. OpenAI API access authorizes OpenAI services, not Montage workspace tools. Keep account-specific auth out of model messages, examples and shared logs, and review what information would cross the hosted execution boundary. If compatible authorization is absent, stop rather than delegating a failing call to a model.
Prefer a direct preliminary diagnostic
For initial validation, use the locally controlled MCP connection's documented lifecycle to inspect tools and select signals_status without initiating an agent run. Discovery alone does not prove account-scoped execution. If your application ultimately uses hosted execution, record that a local success verifies that local path only, not every hosted approval or authorization setting.
Finish with the actual signals_status response
Execute authorized Montage signals_status through the configured direct client and inspect the workspace response. Repair Montage auth if rejected; check Streamable HTTP settings and SDK version if initialization fails. A later hosted check needs its own explicitly approved execution, not an inferred success from this test. No paid Responses call or collection action is necessary for this guide's read-only local verification.
Keep the boundaries clear.
- MCPServerStreamableHttp and HostedMCPTool have different execution and authorization boundaries.
- OpenAI API credentials are not Montage credentials; application-owned authorization is required.
- No paid inference was run, and local validation does not certify hosted execution 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 2026-11-13. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.