Build your first n8n inbox AI agent: a complete walkthrough

What is n8n, and how does it handle AI agents?

How to build AI agents with n8n starts with understanding that n8n is a visual workflow automation tool where AI is just another node in a larger, controlled process. n8n uses a node-based canvas where you drag triggers, AI agents, and app integrations to build end-to-end flows without writing much code.⁵¹¹

n8n’s AI Agent capability sits alongside more traditional nodes like HTTP Request, Gmail, Slack, and Google Sheets.⁵⁹ You chain them into a graph: a trigger (schedule or webhook), an AI “brain”, and downstream actions.

For our inbox-focused tutorial, we’ll use this pattern to build a practical summarize-my-inbox agent that runs on autopilot but still keeps you in control.

---

How to build AI agents with n8n: what are we actually building?

We’ll build a simple but useful AI agent in n8n that fetches email threads, summarizes them, assigns priorities, and delivers a clean daily digest you can scan in under five minutes. The agent will sit in a single workflow: trigger → Gmail → AI Agent / LLM → summary output.

This is deliberately lighter-weight than a full “second brain” and closer to an AI-powered inbox briefing.

Where other tutorials jump straight to advanced multi-agent architectures, here we stay focused on a single agent with a narrow job: turn cluttered emails into structured summaries you can act on.

---

How do you set up n8n for your first AI agent?

To build your first AI agent with n8n, you start by creating an n8n account, opening a blank workflow, and wiring in an AI model credential (OpenAI or Gemini) that your agent will use for reasoning.²³

Step 1: Create your n8n workspace

• Go to n8n and create a free cloud account; you don’t need payment details to start.²
• Once inside, click Workflows → New Workflow to open a blank canvas.²
• Give it a name like Inbox Summary Agent (Daily) so it’s obvious what it does.

If you’re self-hosted, the steps are similar; you just access your own n8n instance instead of n8n’s cloud.⁵

Step 2: Connect your AI model

You have two main beginner-friendly options:

• Google Gemini via API key – n8n’s official “Build your first AI agent” template uses Gemini as the default model and only needs a single API key.³
• OpenAI via API key – equally common in community tutorials and templates in 2025.⁸

For Gemini (mirroring n8n’s own template):

1. Visit Google AI Studio and create an API key in a new project, then copy it.³
2. In n8n, add a Google Gemini or Connect your model node to the canvas.³
3. Open its Credential dropdown, choose + Create New Credential, paste the key, and save.³

For OpenAI:

1. Grab an API key from your OpenAI dashboard.
2. Add an OpenAI Chat Model node.
3. Create an OpenAI credential and paste the key.⁶⁸

At this point you have a “thinky brain” available on the canvas, but it isn’t yet an AI Agent with tools or memory.⁴¹

---

What does the core inbox agent flow look like in n8n?

A practical inbox agent in n8n follows a simple spine: a scheduled trigger runs once per day, pulls recent emails via Gmail, sends them to an AI Agent node for summarization and prioritisation, then delivers the output as a digest.²⁷¹²

The base workflow pattern

The pattern mirrors common email automation examples in 2025–2026:⁴⁷

1. Trigger node – a Cron or Schedule node that runs at, say, 08:30 every weekday.
2. Gmail node – lists or fetches unread emails from a specific label or folder.¹²
3. AI Agent / LLM node – receives email subjects and bodies and produces structured summaries.¹⁹
4. Output node – sends the digest to you via Gmail, Slack, or another channel.⁴

You can extend this with Switch nodes and human-in-the-loop approvals if you want to gate high-priority items or auto-draft responses.²

Why use an AI Agent node instead of a bare LLM node?

n8n differentiates between a simple model call and an Agent that can:

• Accept system prompts that define behaviour.¹
• Use tools like HTTP requests or database lookups.¹⁴
• Store memory across interactions when you use chat triggers.²⁵

For our inbox summarizer, we’ll lean on:

• A clear system prompt.
• Structured JSON output for priorities and tags.

We won’t use long-term memory or external tools yet—that’s the next layer once your base agent is reliable.

---

How do you connect Gmail to n8n for inbox access?

To let your AI agent summarize real emails, you must authorize Gmail within n8n via OAuth, then use a Gmail node to fetch messages from your inbox or a specific label.⁴⁸¹²

Step 1: Create Gmail credentials

n8n’s Gmail docs and 2026 tutorials walk through a standard OAuth flow:⁸¹²

• Add a Gmail node to the canvas.
• In its Credentials dropdown, choose + Create New Credential.
• Select the Gmail credential type and follow the Google OAuth prompts.

Behind the scenes, this creates a client ID and secret and ties them to n8n so the workflow can read and send email on your behalf.⁵⁸

Step 2: Fetch the emails you actually care about

For a focused inbox summary, avoid fetching everything.

Configure the Gmail node to:

• Use List Messages or similar.
• Filter by label (e.g., INBOX, priority-clients).
• Limit the number of messages (say, the last 50 unread) to keep your AI cost predictable.

This “messages” output is what you’ll map into the AI Agent node’s input field.

---

How do you design the AI Agent prompt and output for reliable summaries?

The AI Agent in n8n becomes reliable when you define a strict system prompt and force the model to return predictable JSON containing summaries and priority labels.²⁶¹⁵

Step 1: Add an AI Agent node

On your canvas:

• Add an AI Agent node and connect it after the Gmail node.¹⁹
• In the agent’s settings, select the backing model (Gemini or OpenAI) using the credential you already configured.³⁶

This node will now act as the “brain” that receives email data and emits structured decisions.¹

Step 2: Write a narrow system prompt

Borrowing from real-world agent workflows, you want something like:²⁶⁹

You are an inbox summarization assistant for a solo consultant. You receive a list of email threads with subject, sender, received time, and body. Your job is to:
1) Group related emails into threads.
2) Produce 1–2 sentence summaries per thread.
3) Assign a priority of urgent, today, or later.
4) Suggest a single next action per thread.

Keep it short but specific. The more your agent behaves like a disciplined function, the less you fight vague output.¹⁵

Step 3: Force structured JSON

Downstream nodes work best when the AI output is structured.²⁶

Add a response format instruction such as:

Return a JSON array called threads. Each item must have: subject, sender, summary, priority (urgent|today|later), and next_action.

n8n can then parse this JSON and route based on priority using a Switch node if you choose.²

This pattern mirrors lead-scoring agents where the model returns lead_score and next_steps fields that drive automation.⁶

---

How do you assemble the full “summarize-my-inbox” agent workflow?

You assemble the inbox agent by chaining n8n nodes in order: Schedule trigger → Gmail (list messages) → AI Agent (summarize + prioritise) → Output node (email or Slack digest).²⁴⁷

Step-by-step assembly

1. Schedule Trigger

   • Add a Cron or Schedule node set to weekdays at 08:30.

2. Gmail Fetch

   • Add a Gmail node configured to list unread messages from chosen labels.¹²
   • Map fields like subject, from, date, and snippet or body into a simplified array.

3. AI Agent Node

   • Add the AI Agent node connected to your Gemini or OpenAI model.¹³⁶
   • Paste your system prompt and output specification.
   • Map the email array into a single text or JSON input for the agent.

4. Digest Output

   • Add a Gmail (Send) node or Slack node.
   • Compose a plaintext or Markdown digest using n8n’s expression language, iterating over the threads array the AI returns.

5. Optional: Switch + Human-in-the-loop

   • Add a Switch node based on priority.
   • Route urgent items to a separate channel or a confirmation step before any auto-actions.²

Once wired, click Execute to test the workflow on a small batch of emails before activating it.⁵

Before / after: what changes in your inbox?

State	Before the agent	After the agent
Morning routine	Manually triaging 50+ emails	Reading a 10–15 line digest
Prioritisation	Gut feel, easy to miss urgent threads	Explicit urgent / today / later labels per thread
Action planning	Separate task list in another tool	next_action suggestions embedded in the digest
Time spent	20–30 minutes of triage	3–5 minutes of review

This isn’t a fully autonomous inbox, but it’s a credible first AI agent that saves real time without taking risky decisions.

---

How do you control cost and choose the right model for summarisation?

You keep inbox agent costs under control by limiting email volume, using cheaper summarization-friendly models, and optionally routing all AI calls through a gateway like TrueFoundry for monitoring and switching.¹⁰

Model choice for this agent

For summarisation, you rarely need the most expensive model:

• Gemini Flash / similar “lite” variants are explicitly used in n8n agent workflows for tasks like lead scoring, trading some depth for speed and cost.⁶¹⁰
• GPT‑4o “mini” style tiers via gateways or OpenAI are likewise tuned for high-volume, low-latency tasks.¹⁰

If you use a gateway such as TrueFoundry’s AI Gateway, you can point your AI Agent node at a single endpoint and change the underlying model ID (e.g., openai-main/gpt-4o) without editing the workflow.¹⁰

Practical cost controls

• Limit daily emails to a maximum batch (e.g., 50 messages).
• Strip attachments and signatures before sending text to the model.
• Use short summaries (1–2 sentences) and avoid long quote-backs of original text.

This keeps token usage predictable and aligns with the “small, focused agents” philosophy emerging in 2025 n8n tutorials.⁸¹¹

---

What pitfalls do beginners hit when learning how to build AI agents with n8n?

The main beginner pitfalls when learning how to build AI agents with n8n are vague prompts, unstructured outputs, lack of testing on small batches, and overcomplicated workflows too early.¹¹¹⁵

Community discussions and tutorials from 2024–2026 highlight a few recurring themes:⁹¹¹¹⁵

• Underestimating the learning curve – a “Master n8n in 2 Hours” guide still warns that powerful features introduce complexity.¹¹
• Skipping the AI Agent node – some builders call LLMs directly and miss agent-specific tooling and memory capabilities.¹
• No structured output – free-form text is harder to route and leads to brittle downstream logic.²⁶
• Testing with full inboxes – running on hundreds of emails before validating outputs makes debugging painful.¹⁵

A calmer, more sustainable approach:

• Start with one focused agent (like this inbox summarizer).
• Test with 5–10 messages at a time.
• Iterate prompts and JSON shapes until the digest looks consistent.

From there, you can introduce more tools—like calendar lookups, CRM updates, or Slack alerts—one node at a time.²³⁷

---

Where can you go next after this inbox agent?

Once your inbox agent is stable, you can extend it into richer AI workflows by adding tools, memory, and multi-step routing that still sit inside n8n’s visual canvas.¹⁴⁹

Some natural next steps:

• Add a “reply draft” path – for urgent threads, have the agent propose a draft response, then send it to you for approval via Slack.²⁷
• Log summaries to a knowledge base – insert each thread’s summary into Notion or Google Sheets to build an audit trail.⁵
• Introduce conversation memory – if you later expose the agent via an AI Chat Trigger, add a Memory node so it remembers previous context across messages.²⁴

n8n’s AI integrations page lists 400+ apps you can connect to your agent, from CRMs to helpdesks, turning your inbox summarizer into a hub that kicks off other workflows based on email content.⁵⁹

The point is not to chase “full autonomy” but to ship a series of reliable, narrow agents that remove tedious work for you—and this inbox digest is one of the safest, highest-leverage places to start.