How to Build a No-Code Personal Assistant AI Agent in n8n

TL;DR

A personal assistant AI agent built in n8n connects to Telegram and uses GPT-4o as its brain, letting you request tasks by text and get results instantly, 24/7, for a fraction of a human assistant's cost. It calls custom sub-tools to get and send emails, manage calendar events, and update a project database.

Transcript

I'm super excited to share this AI agent build with you guys today I definitely think that it's going to change the way that I do my business especially anything personal Administration related and that's because I'm going to consider this a human personal assistant except for unlike a human this thing is going to be available to me 24/7 it's going... Read More

Key Insights

• The agent is controlled entirely through Telegram, so requests can be typed from a phone or desktop and the agent either returns a result or reports that it completed a back-end action like updating a database or setting a calendar event.
• GPT-4o serves as the brain of the agent entity, providing the reasoning that decides which sub-tool to call to complete a user's request.
• A Pinecone vector store holds contact information, letting the agent look up a person like Jenny or Mike by name and retrieve their email address before sending a message.
• Custom sub-tools built inside n8n (get emails, send email, get calendar, set calendar, update database, summarize database) are marked with a logo showing they connect to workflows the creator built in the same software.
• Email summarization works by filtering Gmail for inbox messages marked unread and received after the requested date, so already-read emails from the same day are excluded from the summary.
• An OpenAI node interprets relative dates by receiving the current date via a JavaScript '$now' script plus examples for 'today' and 'yesterday', ensuring consistent date parsing since the model does not always infer date math reliably.
• An aggregate node combines multiple email items into one output so Telegram returns all summaries at once instead of only messaging back the first email.
• Each sub-tool workflow must write a field named 'response' (or a value like 'sent' or 'done') that signals back to the agent that the task completed and supplies any information to relay to the user.

Questions & Answers

Q: How do you interact with the n8n personal assistant AI agent?

You interact with the agent through Telegram, which can be pulled up on a phone or desktop. You simply type a request asking it to do something, such as summarizing emails or sending a message. The agent runs through its logic and either outputs a result to you or tells you it performed a back-end action, like updating the database or setting a calendar event, then reports back on Telegram.

Q: What model powers the AI agent's reasoning?

The agent uses the OpenAI chat model GPT-4o, described as the brain given to the actual agent entity itself. This model provides the core reasoning that lets the agent decide which sub-tool to call to complete a request. The individual sub-tool workflows also use OpenAI nodes internally for tasks like parsing dates and summarizing email content before returning results to the agent.

Q: What is the Pinecone vector store used for in this agent?

Pinecone is the vector store database connected to the agent, and in this use case it holds contact information. This lets you say something like 'can you email Mike' or 'can you email Jenny,' and the agent will know how to go find that person, retrieve their email address, and actually send an email. It requires connecting embeddings to the tool to function correctly.

Q: How does the agent decide which emails to summarize?

The Gmail node filters for emails in the inbox that are marked as unread and received after the requested date, such as October 11th. Because the filter targets unread messages, already-read emails from the same day are excluded and not returned. In the demo, two unread emails came through and were summarized, while a read email from the same day was skipped.

Q: How does the agent handle relative dates like 'today' or 'yesterday'?

An OpenAI node is prompted to identify what date the user is asking for based on the JSON query and the current date. The current date is supplied through a JavaScript script using '$now,' which returns the current date-time. The prompt includes two examples: if the query asks for today, return today's date; if it asks for yesterday, return the day before. These examples ensure consistent parsing since OpenAI does not always know the correct date math.

Q: Why is an aggregate node used in the email summary workflow?

The aggregate node combines multiple email items into a single output. Without it, the workflow would only message back the first email on Telegram. By aggregating, if there were eight emails needing summarization, the agent would be able to send all eight of them in one message rather than just one. It turns two items into one before the response is sent back.

Q: How does a sub-tool signal to the agent that it finished a task?

Each sub-tool workflow sets a field named 'response' with the content or a status value. For the get-emails tool, the response field carries the summarized content the agent relays back. For the send-email tool, an edit field set node manually creates a 'response' field set to 'sent' (or 'done'), which indicates to the agent that the workflow completed successfully so it can report back on Telegram.

Q: What does the send email workflow look like in n8n?

The send email workflow is only four nodes and is fairly simple. Incoming information includes who the email is going to, the subject, and the message. An OpenAI node parses this into the correct parameters, outputting the message, recipient, and subject. A Gmail node then sends it, with the send-to, subject, and message fields dragged in and the email type set to text. Finally, a response field set to 'sent' communicates completion back to the agent.

Summary & Key Takeaways

• The build is a personal assistant AI agent created with no code in n8n, interacted with through Telegram so the user can ask it to do tasks from a phone or desktop. It is pitched as available 24/7, responding immediately at a fraction of the cost of hiring a human personal assistant.

• The agent uses GPT-4o as its brain and connects to a Pinecone vector store holding contact information, so it can find a person's email by name. Custom n8n sub-tools handle getting and sending emails, getting and setting calendar events, and updating and summarizing a database.

• In the email demo, typing 'summarize my emails from today' filters unread inbox emails after the requested date, uses an OpenAI node to parse the date and another to summarize, aggregates them into one message, and returns a 'response' field to the agent to relay back on Telegram.