Google’s Connected Apps Platform: MCP Integration Still Needs a Submission Boundary
Google’s Connected Apps Platform brings MCP-backed apps to Search and Gemini. The hard part is still testing, review, identity, and rollback.
When I build a product, it is tempting to treat “the API call works” as “the agent integration is ready.” The uncomfortable work starts later. Which records can the tool see? Does the user understand the authorization? Who can stop a bad write? On October 28, Google announced a preview of the Connected Apps Platform. Developers can connect a cloud-hosted MCP server or Android MCP through AppFunctions APIs, register and test the app in the Connected Console, then choose whether to submit it for Search, the Gemini app, or both.
An MCP server is not a blank cheque
An MCP server might expose a dozen tools. I would not hand all of them to a model by default. Split actions by risk: reading a calendar is low risk, creating an event needs more care, and cancelling an event or sending a message is high risk. Tool descriptions should state inputs, permissions, and side effects. The server must validate them again; a model’s JSON is not an authorization decision.
A calendar write should have an explicit boundary:
type CreateEventRequest = {
title: string;
start: string;
end: string;
confirmationToken: string;
};
if (!isValidConfirmationToken(req.confirmationToken)) {
throw new Error("human confirmation required");
}
The token separates a suggested action from an approved action. Google says the platform supports registration, testing, submission, and tracking for traffic, latency, and errors. Those operational metrics are useful, but they do not replace confirmation for a specific write. A low average latency can still accompany a correctly formatted action against the wrong account.
Identity needs more than a verified icon
The announcement describes controls for brand identity, including verified logos and authenticated accounts. That helps users understand who is connected, but I would not confuse identity with authorization. Every request still needs a server-side binding to the user, tenant, and resource scope.
My review list would include questions like these:
- Does the request carry an explicit tenant ID that the model cannot rewrite?
- Can a read response leak another user’s data, internal prompts, or secrets?
- Does every write need a one-time human confirmation by default?
- Can the server behave idempotently after a timeout or duplicate delivery?
This is why I dislike the idea of one universal tool. It saves a few lines of model-generated calls while making the review surface huge. I would rather expose list_projects, get_project, and request_deploy than give the model an admin_api that can do anything.
Submission must include a rollback path
Putting registration, testing, and submission into one platform is useful for a small team. The important part is not getting the first submission accepted; it is being able to withdraw safely. Before launch I would prepare a kill switch that makes the MCP server return “human handling required” for writes, keep the last known-good version, and store every tool call in our own audit table instead of relying only on aggregate platform metrics.
If the app appears in both Search and Gemini, I would test the contexts separately. The same tool may receive different intent, login state, and visible data at each entry point. A successful demo in one surface does not prove safety in the other. Tests should cover a signed-out user, an expired authorization, a double click, and a permission revoked immediately before the write.
I expect Connected Apps to move MCP from developer desktops into higher-intent user entry points. That raises the product responsibility as well as the reach. The connection button is the beginning; review, confirmation, and rollback will decide whether an integration can be maintained for years.
I would test the consent copy as part of the product surface, not leave it to default wording. A user should know which account will be read, what the action changes, and whether it can be undone. Store the resource ID and authorization version in the audit record so an incident can answer what was permitted at that moment. If the tool scope expands after approval, I would submit it again rather than treating the old review as a permanent pass.
For payments, bookings, or data deletion, I would make preview and commit separate tools. Preview returns structured information; commit requires a fresh confirmation token and reads current state again on the server. If the price, time, or inventory changed before confirmation, the old approval is rejected instead of being applied to stale data.
These details also affect support cost. Users may never know what MCP is, but they will ask why a calendar event appeared twice or why a payment did not happen. The product needs an audit trail that a person can understand.
Google Connected Apps Platform announcement, Connected Apps developer documentation, Google’s MCP developer resources
