If a Google Workspace MCP connection is being rejected, first identify the endpoint and OAuth flow. Do not start by deleting a working client, adding broader scopes or assuming every authorization error has the same cause.
Google's agentic-client requirement is a real diagnostic lead, but it is not a universal explanation for Gmail, Calendar or Drive integration failures. This guide separates the dated announcement from the current setup instructions and proposes a cautious recovery checklist. It does not report a Nerova incident, integration test or configuration change.
What changed, and when?
Google's October 12, 2026 Workspace release notes announced that MCP servers would reject unmarked clients starting November 11. They also described a November 12 boundary after which an unmarked client would need replacement. REST-only integrations were excluded from this change. Official dated announcement.
Do not turn that announcement into an instruction to replace every client today. Current setup documentation gives conditional recovery branches. Check what your actual client supports and which authentication flow it uses before changing it.
The current client-type decision tree
The current Gmail MCP troubleshooting documentation distinguishes these cases. Gmail MCP remains a Developer Preview; confirm your eligibility and product-specific setup requirements.
| Client or flow | Documented route |
|---|---|
| User-delegated OAuth client; setting available | Select Use this client for an AI-powered agent in Google Auth Platform > Clients. |
| User-delegated client; setting unavailable | Create an agentic OAuth client and update the integration to use it. |
| One client serving agent and non-agent features | Separate the agent client; retain the non-agent client. |
| Service account with domain-wide delegation | No checkbox applies. Missing it does not cause this rejection; administrator delegation controls and sensitive-action confirmations still apply. |
| Apps Script managed default client | For MCP, use a separate agentic client with an OAuth2-library token, not ScriptApp.getOAuthToken. A standard Cloud project alone does not change this. |
These branches are not permission to move to domain-wide delegation as a workaround. Ordinary Apps Script that does not call Workspace MCP is not affected by this requirement.
Do not confuse a client label with authorization
The agentic setting is a self-declaration. It grants no additional scopes or privileges and does not remove user-consent or app-verification requirements. Google directs mixed-use applications to maintain separate agent and standard clients. Client secrets belong in secure secret storage, not source control or a troubleshooting screenshot. Google's client-management guidance.
Keep four questions separate: which application is asking, which user or service identity is represented, what access was granted, and what action is permitted now. A successful sign-in answers only part of that chain. A tool being listed does not establish that a particular document is readable or that a sensitive action is authorized.
Capture the smallest useful diagnostic record
Before editing configuration, record enough evidence to distinguish causes without collecting private content:
- The affected product and exact service hostname, with personal identifiers removed from paths and queries.
- Whether the request targets Google's native Workspace MCP service, an ordinary REST API, or a third-party MCP wrapper.
- The authentication-flow class and an internal configuration reference, not a token or secret.
- The time, HTTP status and redacted error description, plus whether failure occurs during sign-in, tool discovery or a specific operation.
- The last known successful version and the most recent relevant configuration change.
A generic 401 or 403 is a starting observation, not a diagnosis. Check the product's current documentation and the actual error before choosing a repair. If several unrelated clients fail together, investigate their shared dependency rather than independently rotating all of them.
A controlled recovery sequence
- Confirm ownership. Identify who may administer the client and approve access changes. Do not borrow a broader credential merely to see whether the error disappears.
- Inventory dependencies. Identify the applications and non-agent features that currently use the client. Preserve a record of the existing approved configuration without exporting secret values into the incident notes.
- Choose the matching branch. Use the current client-type guidance above. If the console and documentation disagree, collect redacted evidence and escalate through the appropriate administrator or support route rather than guessing.
- Test the smallest authorized operation. Use an approved test account or tenant with a synthetic fixture. Verify a permitted read-only operation against that fixture before attempting an explicitly approved write.
- Verify the result at the destination. Record the returned object or state using only the minimal non-sensitive test evidence. Authentication success alone is not recovery of the business workflow.
- Check unaffected paths. Re-test the standard feature that shares the surrounding application. A repaired agent path should not silently break an existing user-facing workflow.
- Roll out with an owner and a stop condition. Observe failures and unexpected access. Retire old credentials only through the owner's approved process after dependencies are understood, not as an automatic first step.
This is a proposed operational sequence, not evidence that a particular deployment has passed. A failed test should retain its real outcome. Do not classify a manual workaround as an automated recovery.
Keep security controls in the recovery path
Google requires screening prompts and responses for malicious content or prompt injection, using Model Armor or a documented alternative that enables users to accept the risk. Its security guide warns that Model Armor logging can capture entire payloads and that routing behavior can affect data residency. Review those choices deliberately; do not switch off protection or turn on full-payload logging as a generic fix. Workspace MCP security guidance.
Treat content returned by a mailbox or document as data, not new permission. For testing, a fictional document can include an instruction-like sentence that the application must not treat as authority. Keep tool execution constrained outside the model, and check that human approval requirements still work after the connection is restored.
What counts as fixed?
Define recovery as the intended, authorized operation succeeding with the correct identity and bounded access, while unaffected features still work. Separately record sign-in failures, tool-discovery failures, access-policy denials and task failures. Combining them into one “agent success” number hides what needs attention.
For example, ten successful sign-ins followed by two successful authorized task reads are not ten completed tasks. Report the numerator, denominator and test conditions. Do not infer production reliability from a single test account or a brief green dashboard.
For the broader execution contract, see tool calling and permission boundaries. If a recovered workflow repeats actions, diagnose agent loops and unsafe retries separately rather than changing OAuth again without evidence.