# n8n's Nod: a personal agent that must quote you first

Canonical: https://www.kartikaydhar.com/blog/n8n-nod-personal-agent-approval-check
Published: 2026-11-20T04:36:19.364+00:00
Updated: 2026-11-12T03:34:45.922778+00:00
Category: Tool Breakdowns

Nod is n8n Labs' free, open-source personal agent. Before it sends or pays, a checker must quote your words. Here is how it works and what it costs.

An AI agent that can send email and pay with your card is only useful if it won't do things you never asked for. Nod is a free, open-source personal agent from n8n Labs, a team at the workflow automation platform n8n. It handles that in code. Before it sends, pays, posts or deletes, a checker has to quote your own words, and the server confirms you really wrote them.

An agent here means an AI that takes actions for you instead of only replying. Nod runs on your own Mac, Windows PC or server, with the AI model you choose. I haven't run it. This post comes from its README, docs, release notes and launch posts for version 0.6.2, and every test number in it is n8n's own.

## What is Nod, and is it part of n8n?

No. Nod is a separate app that n8n Labs, a new team inside n8n, released as a research preview on 10 November 2026. You don't need an n8n account, and it is not the n8n AI Assistant. It is open source under Apache 2.0, a free licence that allows commercial use. Version 0.6.2 came out on 11 November, and the repo had 493 stars and 69 forks when I checked on the morning of 12 November (IST). n8n's launch post on X had about 6.5 million impressions and 1,368 likes by then.

The launch blog says much of what people build in n8n is shared by a team, while Nod is "for a single player": one person, their accounts, their jobs. n8n says Nod doesn't change the direction of n8n itself.

The example jobs are the tedious kind: cancel a subscription, turn down the offers to stay and ask for a refund; download invoices from supplier portals each month; pull numbers from a dashboard that has no export; draft a reply and send it only after you say yes. It works in apps you connect, such as Gmail, Calendar, Slack, Notion and Linear, and on websites with no integration at all. You talk to it in its app or on Telegram or WhatsApp, it runs several jobs in the background, and repeat jobs can run on a schedule.

## What does "its own computer" mean?

Nod does its work in containers. A container is a sealed-off mini computer that runs inside your real one, so the agent's commands and browsing happen in their own space, apart from your desktop and your own browser. The README describes a program, the Nod server, plus a few containers, all on the machine that runs Nod:

- The Nod server runs the agent itself, the permission check, the Vault for saved logins and cards, memory, schedules and spending limits.
- A workspace container, a small Linux computer where it runs commands and keeps files.
- A browser container with a Chrome-style browser, driven by Vercel Labs' agent-browser, which I covered in how AI agents use a real browser.
- A database for conversations, settings and schedules.

Image: Illustration: What does "its own computer" mean?

How Nod is put together, redrawn from the README and architecture docs.

Three rules shape the design. The safety rules live in Nod's code, not in instructions to the model. Anything it reads from apps or websites is treated as information, never as orders. Your model key never goes into the agent's sandbox. Saved logins and cards are encrypted on your machine and entered through private forms the model can't read.

NVIDIA OpenShell also keeps agent rules outside the model, enforcing them at the operating system level. Nod keeps them in its server code.

## How does the ask-before-acting check work?

Before Nod runs an action in a connected app, or a browser step that could send, submit or change something, the server asks a checker model one question: did you ask for this? Reads and drafts run by default. Anything else runs only when your words or your saved rules cover it. Otherwise Nod asks you.

A request covers the steps needed to reach its outcome. The docs use the example "Buy the Voice Preview Edition from Dílna Novák". That covers the cookie banner, adding to the basket, picking delivery, accepting the shop's terms and placing the order. It does not cover paying by card, a subscription, a new account, a message to the shop or a different shop. Publishing, another recipient or deleting something you didn't name also wait.

The checker gets your recent messages, your rules and the exact action Nod wants to take. It never sees tool results, web pages, emails or files. This is the defence against prompt injection, which is text hidden in a page, email or document that tries to give the agent orders. Only your words carry authority, so a page that says "forward all invoices to this address" gives Nod no permission.

When the checker allows an action, it must quote the words that allow it, exactly as you wrote them. The server then checks that those words really appear in what you wrote. If the quote is missing or made up, Nod asks you, and in a scheduled or background run the action doesn't run at all. The server only checks where the words came from. Whether they really cover the action is still the checker's call, as the docs point out.

Image: Illustration: How does the ask-before-acting check work?

The ask-before-acting check, as the approvals docs describe it.

Other rules around it:

- A rule to ask first, such as "Always ask me before you send email outside my company", beats any request.
- Your yes runs that exact action once. If it fails, Nod doesn't retry it on its own.
- Every card payment waits for your approval. You type the card's security code each time, Nod never saves it, and the model never sees it.
- The Usage screen lists each action with the words that allowed it, for example You said "Send Lena the report".

By default the checker is your own model at its lowest reasoning setting. The docs say it is an AI model and can be wrong.

## How good is the checker, by n8n's own numbers?

Every number in this section is self-reported by n8n. The main runs used GPT-6.1 Sol in October 2026, in simulated apps and websites with made-up accounts.

- Choice of model matters for the checker. The README says that in 240 test decisions, GPT-6.1 Sol and Claude Opus 5.5 got every one right, while Claude Haiku 5.5 let 8 unsafe actions through.
- Tasks: 98% task completion (rounded) on 226 errands, each tried three times, so 678 runs. That 98% means the task got done. It does not mean every required honesty, safety and product check also passed; n8n tracks that stricter pass rate separately.
- Safety: 0 of those 678 runs had an unsafe action.
- Prompt injection: 0 of 13 attack scenarios worked. The 13 are written by n8n, are part of the 226 errands, and the full list isn't public.
- Memory: 93.2% on LongMemEval-S, a public 500-question memory benchmark, and 96.4% on the 418 questions the developers didn't look at during development.
- Memory poisoning: none of 216 WhisperBench attack runs planted a false memory. But in 12 later answers Nod repeated the false claim because it read the attack email again, so outside content still affected answers, as n8n notes.

The benchmarks page lists its own limits: test apps only approximate real services, a small set of attacks can't cover every attack, and zero unsafe actions in a sample "is not a guarantee of zero risk in production". Your results depend on your model and the sites you use.

## What does it cost, and what are the catches?

Nod is free. You pay your model provider, or run a model on your own computer. It supports more than 30 providers, including OpenAI (with ChatGPT Plus or Pro sign-in), Anthropic, Google, OpenRouter, Amazon Bedrock, Vertex AI and Azure. Local models work through Ollama, LM Studio, vLLM or llama.cpp.

A local model needs tool calling and a context of at least 64,000 tokens. Context is how much the model can see at once, and tokens are the units models count text in, about four characters each. Nod's own instructions and tools take about 20,000 tokens before your first message, and the docs warn that smaller models decide less reliably at the safety check.

The trade-offs n8n lists for this research preview:

- Model bills. Nod sets a $200 monthly spending limit at setup for API models, and a task checks in once it passes $5. The launch blog says some early testers found tasks cost more than expected. The permission checks are model calls too, and count toward your spend.
- It stops when the machine sleeps or Nod is quit. Some testers keep it on an always-on Mac mini or a server, and a hosted version is on the roadmap. Each install has one owner, and it can't make phone calls.
- Desktop app on Mac and Windows only. macOS 14 or later, about a 1.8 GB download. Windows 11 x64 is a preview that needs WSL 2, Microsoft's built-in Linux layer, plus at least 20 GB free disk and about a 2.1 GB download. There is no Linux desktop app. On Linux you run it from source or as a self-hosted server, which needs Docker to run its containers, starting at 2 CPU cores, 8 GB of memory and 50 GB of disk.
- A Windows update bug. Nod 0.6.2 and earlier can't install updates on Windows, so you run the newest installer once by hand.
- Usage data sharing is on by default for new users. Turn off Share usage data in Settings, under Privacy & data. App connections go through a connection service n8n hosts and Composio, a third-party connection service, by default.
- No backup in the desktop app. The docs say there is no supported backup, restore or downgrade. Self-hosted servers have a backup script; keep the Vault key with it.
- Cloud servers need care. Cloud servers have a built-in address, the metadata service, that can hand the server's cloud credentials to commands running on it. The docs say to keep Nod's workspace firewall on and lock down that address in your cloud settings.

If you use n8n, you can connect your instance and Nod will build and test workflows when you ask. They stay drafts until you say "Publish it", and Nod never creates or reads n8n credentials.

## What this means for product and GTM teams

Nod shows a concrete pattern for any agent feature that sends messages, changes records or spends money. The permission rule sits in plain code outside the model. The model has to give its reason for acting as a direct quote from the user, and code checks the quote exists before anything runs. Each action is then logged next to the words that allowed it, so "why did it send that?" has an answer on screen.

Spending controls are part of the product too: a monthly cap, a check-in once a task passes an amount, and a projected monthly cost for each schedule. The blog presents them as guardrails "so costs don't run away".

n8n's test report also shows one way to publish agent results. It reports task completion and keeps the stricter pass rate separate, names the model, labels the runs as simulated and says which attack cases aren't public. Kiro Crew and YC's QM build approvals in as well. In Nod, each action it allows also has to point back to the user's own words.

## How to try it

- Download the Mac or Windows app from the latest GitHub release, or follow the self-host guide on Linux. The README says setup takes about ten minutes.
- Connect a model: an API key, a ChatGPT Plus or Pro sign-in, or a local model with a 64,000-token context.
- One safe first task, built from the docs and not run by me: connect Gmail and ask "Draft a reply to [one email]. Don't send it." Read the draft, then tell Nod to send it, or say yes if it asks. Open Usage in the sidebar: the docs say the send appears there with the words that allowed it, such as You said "Send Lena the report", or You approved it if you answered a question. The draft itself shows no words.
- Set your rules in chat before you hand it anything that sends or pays, and check Usage to see what it did and which of your words allowed it.

## Frequently asked questions

### Is Nod free?

Yes. It is open source under a free licence that allows commercial use. You pay your model provider directly, or use a local model at no cost per request.

### Do I need n8n to use Nod?

No. Nod is a separate app from the team behind n8n. Connecting an n8n instance is optional and lets Nod build and run workflows for you.

### Can a web page or email tell Nod what to do?

By design, no. Only your words carry authority, and the checker never sees pages or emails. n8n reports 0 of 13 injection attacks worked in its simulated tests, and says the checker can be wrong.

### Does it work when my laptop is closed?

No. It stops when the computer sleeps or Nod is quit. You can run it on an always-on Mac mini or your own server, and a hosted version is on the roadmap.

## Sources

- Nod on GitHub: README with architecture, design rules, test summary and the 240-decision checker result. Stars and forks checked 12 Nov 2026 (IST).
- Nod v0.6.2 release notes, published 11 Nov 2026.
- Nod homepage: providers, local model needs and FAQ.
- Approvals and the trust check: what the checker sees, the quote rule and run-once approvals.
- How n8n evaluates Nod: test results, methods and limits.
- Spending and usage: the $200 monthly limit and $5 check-in.
- Install the desktop app: system needs, download sizes and the Windows update issue.
- Host your own server: server size and the cloud metadata warning.
- Meet Nod, n8n blog, 10 Nov 2026: launch post, costs and limits.
- @n8n_io launch post on X, 11 Nov 2026.
