Build a no-code Gemini email assistant in n8n

Define the assistant’s job and a small test

The June 2025 lesson presents an n8n agent that can read, draft and send Gmail messages, with memory and a system prompt for context. It contrasts an agent using tools with an ordinary conversational answer. The useful first result is a checked response to a known message, not unrestricted control of an inbox.

1. Open https://n8n.partnerlinks.io/trxvmi48ria4 and follow the workflow overview at 5:46 and build at 8:59. Use a test account or clearly identified messages and define the task the assistant should perform.
2. Write the intended tone, information sources and review point into the instructions. Keep a request to analyse or draft separate from permission to send so the test has a visible boundary.

Connect Gemini and inspect Gmail access

The model connection appears at 13:12 and Gmail send/get tools at 18:29. Authentication, correct message selection and a useful response are different checkpoints. The model’s ability to generate a plausible reply does not establish that it read the right email or preserved its details.

1. Follow the demonstrated model and Gmail credential route and record which account and model are actually selected. Ask the agent to retrieve one known message and compare sender, subject and content with the original.
2. Generate a reply as a draft and inspect the recipient, claims and promised actions. Test a request where information is missing and check that the assistant leaves the gap visible rather than inventing a commitment.

Add memory and verify the external result

Memory and tools are expanded at 22:54. Saved context can make follow-ups more useful, while facts from one email should not silently become instructions for every future message. Keep durable preferences distinct from temporary correspondence and recheck the target thread.

1. Test a follow-up that uses an approved preference and inspect how it affects the draft. Change the target message and confirm that unrelated details from the previous test are not reused.
2. For an explicitly approved test send, verify the complete final draft first and inspect the actual sent message afterwards. Record failures and model usage separately from the successful generation of text.

The takeaway

Test retrieval, drafting, memory and sending independently. A useful email agent preserves the exact message context and makes external action reviewable.