Nod From n8n Labs: Great Demo, 14 Minute Wait on Day One

Emer McArthur posted on the n8n forum on launch day, excited to get Nod set up. Then the Claude Premium plan wouldn't connect, because Nod wanted an API key and credits instead.

So the free Gemini tier got a turn. The first question took 14 minutes to answer.

Fourteen minutes. That is longer than doing the errand yourself.

I read that post three times. Not because it's dramatic. Because it's the exact shape of every agent launch i've watched this year.

The demo is a browser opening tabs at triple speed. Day one is someone staring at a spinner, wondering if they broke the config.

Nod is worth writing about anyway. It shipped on November 10 from n8n Labs, the experimental arm inside n8n.

Apache 2.0 license. Runs on your Mac or Windows PC. It brings its own sandbox and its own Chromium.

You hand it a job. It goes and does the job on real websites.

That's the whole pitch, and it's also where things get messy.

What it actually is

Most AI tools with a browser are a chat window with a plugin bolted on. Nod is not that. The app spins up a Linux sandbox on your machine, runs a Chromium inside it, and lets the agent click around like a person would.

The stack is boring in a good way. One Node process, a few Docker containers, PostgreSQL, and the Pi agent runtime. Everything sits on the box you installed it on.

That matters because of what it lets Nod do. It can log into a supplier portal that has no API. It can read a dashboard with no export button and type the numbers into a spreadsheet.

It can also sit on a checkout page and wait for you to say go.

I've used tools that claim this. Usually they mean "we have a screenshot tool and a lot of confidence."

Nod actually drives the browser. Whether it drives it well depends on which model you point at it.

And the model choice is not a small setting. Point it at a cheap model and watch it fumble a login form for six minutes. Point it at a good one and your token bill shows up before your invoice does.

The 98% problem

The headline number is 98% task completion across 226 errands, run three times each. Zero unsafe actions in 678 runs. Memory recall at 93.2% on LongMemEval.

Those numbers are good. They also come with a footnote the page does include: GPT-6.1 Sol, October 2026, in simulated accounts and websites.

Simulated. That's the word to sit with.

A simulated supplier portal doesn't have a cookie banner covering the submit button. A simulated bank doesn't log you out mid-flow because your session looked weird. A simulated checkout doesn't throw a 3D Secure prompt at you and wait.

I'm not saying the benchmark is fake. I'm saying the gap between 98% in a test setup and 98% on a Tuesday afternoon is where products die.

The other thing to know is who Nod is for. n8n's own framing is sharper than most launch posts:

That's a clean split, and it's the reason Nod is a separate app instead of a new tab inside n8n. You don't need an n8n account. If you have one, you can connect it and Nod will build and test workflows there.

It remembers, and you can check

Here's a question people always ask about personal agents. Will it remember that my accountant is Anna and not Bob?

Nod's answer is a visible memory store. Every note shows the message it came from. You can read it, edit it, or delete it.

It also learns your email style from sent mail, with names and numbers stripped out first.

The 93.2% recall number is the one i believe most, because memory is a solved-ish problem next to browser automation. Retrieval is easy. Knowing which fact is still true is hard.

Nod handles that with a tell-it-once rule. Say something changed and it uses the new version from then on.

Small thing. But it's the difference between an assistant and a stranger you re-brief every morning.

The laptop dies, the job dies

Most tutorials tell you to install it on your laptop and forget about it. That works right up until you close the lid.

When the machine sleeps, Nod stops. Scheduled jobs stop with it. The morning brief doesn't arrive because the laptop was in a bag on a train.

n8n knows this. Their blog says some testers keep it on an always-on Mac mini or a server, and a hosted version is on the roadmap.

The comment that stuck with me came from a Redditor who tried the always-on route. Their Mac rebooted for an update, sat at the login screen, and every scheduled job stayed dead for eight days before anyone noticed.

Eight days. No error. No alert. Just silence.

So the honest setup looks like this:

1. Install on the laptop to try it.
2. Move it to a box that never sleeps.
3. Add a check that texts you when a job didn't run.

That's three pieces of infrastructure for a tool that's supposed to remove work from your week.

Local is not offline

"Local" is the word doing the most work in the launch posts. Your files, memory, and saved logins stay on your machine. True.

But local isn't offline. Your chosen model provider gets the conversation, the pages Nod reads, and screenshots. App connections route through a connection service that n8n hosts by default.

Usage analytics are on by default, and the docs are unusually clear that a fresh install can fire an event before you even reach the privacy switch.

I like that the privacy page exists and reads like a person wrote it. It shows you a scrubbed trace where "Jordi" is still sitting there in the text, because name detection by shape doesn't catch every name.

That kind of admission builds trust. Most vendors would have deleted that paragraph.

And the permission check is itself a model. n8n says so plainly. It can make mistakes, which is why anything your words don't cover waits for your yes.

Small tangent. I have a friend who self-hosts everything. His photos, his notes, his calendar, a wiki he keeps for his own recipes.

He once spent a Saturday migrating a to-do app because the maintainer changed the license.

I asked him why. He said he didn't want to wake up one day and find his stuff gone.

I used to think that was paranoid. Now i keep a folder called exports where i dump a copy of everything every few months, and i feel calmer for it.

The catch is that self-hosting is a hobby dressed as a setting. Every service you run is a thing that can break at 2am while you sleep.

He knows this. He has alerts.

Most people don't want alerts. They want the errand done.

And that is the honest pitch for Nod. Not automation. A list on your phone that finally shrinks.

Mine has eleven items on it and the oldest one is from last December.

The real talk

Most people don't need this. If your week has no recurring logins and forms, Nod is an elaborate way to send an email.

If you're not comfortable pasting an API key into a settings screen, wait. Setup is about ten minutes when nothing goes wrong, and day one suggests things go wrong.

If a job has to run at 6am while your laptop is shut, this is not the product yet. Use a hosted agent or a plain n8n workflow on a server.

And if money is tight, watch the meter. n8n admits early testers found tasks cost more than they expected. The $5 check-in exists because someone got surprised.

The software is free. The errands are not.

What i keep coming back to

The 14 minute Gemini reply is the detail i can't shake. Not as a knock on Nod. As a reminder of what day one actually feels like.

You read the launch post. You picture the version where the browser opens, the invoice downloads, the refund lands. Then you spend your first evening debugging model access instead.

That gap closes or it doesn't. n8n is calling this a research preview and telling people to expect rough edges, which is more honest than most launch copy.

I'll keep my install. The list is still there. The oldest item is still from last December.