Building a calendar and email assistant, then getting stuck on permissions and schemas

Check my calendar, find an email, and turn it into a task if needed. I wanted to handle some of the things I normally do across several screens through a conversation, so I started building a personal AI assistant.

In July 2026, I integrated tools for Google Calendar, Tasks, Drive, Gmail, Contacts, Docs, and Sheets into a setup using n8n and Discord. The difficult parts were often outside the conversation itself: permissions and the API calls behind it.

Adding a scope did not update existing consent

An OAuth scope describes the operations an application is allowed to request. I added scopes to the application configuration, but credentials obtained under an earlier grant did not automatically gain those permissions.

The code listed the permission I needed. The API still refused the operation. Looking only at the configuration made this easy to miss.

I needed to obtain consent for the expanded scope. Since then, I have treated the permissions an implementation requests and the permissions actually granted as separate things to check. Being able to sign in is not proof that a particular API operation will work.

One incomplete array schema stopped the request

Another failure came from a tool definition sent to Gemini. I had declared an ARRAY but omitted items, which describes the type of its elements.

The request containing the tool definitions failed with a 400 before the tool could run. Because several tools were bundled into that request, it looked as though the whole conversation had stopped working.

Changing the prompt would not repair a malformed schema. I inspected the definitions and supplied the element type. As the number of tools grows, testing only through conversation makes this kind of failure harder to isolate.

Test the connection separately from the conversation

I moved the Google API integration from direct fetch calls to googleapis and added a connection-check script. Testing authentication and API calls on their own helps distinguish a connection problem from a problem in the conversational layer.

Reading an email and sending one also have different consequences when something goes wrong. For actions that affect another person, I want to keep a human decision about what will be sent.

This article looks back at the July implementation. In September, I moved the development environment from Windows to a Mac; reconnecting services and verifying startup after that move remained separate work. Having written the tools is not the same as having them reliably available in my everyday environment.

Related: Building with AI