SIGNALS is Earthchain's monthly newsletter for sustainability teams, sharing practical lessons on using data and automation to support better decisions and real action.
Each edition is planned and written in a Google Sheet. I wanted Claude to draft and edit editions directly, without giving an AI agent the keys to the whole Google account. That account also holds a CRM sheet with customer names, emails and phone numbers.
The approach
Instead of a general-purpose Sheets integration, I built a small Model Context Protocol server that is deliberately narrow:
- It's bound to a single, hard-coded spreadsheet ID. No tool accepts an ID as input.
- It holds only the Sheets scope, not Drive, so it can't list or open other files.
- Its tools are shaped around the editorial job (list editions, read an edition, write a field, lint copy against the house style), not around raw cell access.
Why it matters
This is the pattern I recommend to teams adopting AI agents: design the tool surface around the task, and make the dangerous thing impossible rather than merely discouraged.