FR
live
AI

n8n Labs launches Nod, the open-source personal agent that works on your computer

In November 2026 n8n Labs released Nod as a research preview: a standalone personal agent, open source under Apache 2.0, that runs in a local virtual machine and asks permission before sending, paying, or deleting. Solo operators and freelancers get an agent whose trust boundary is explicit — worth testing while keeping its limits in mind.

A stack of filled-in paper forms, a single amber approval stamp pressed onto the top page.

November 2026. n8n Labs, the new team within the maker of the n8n automation platform, releases Nod as a research preview: a standalone personal agent that installs on Mac or Windows and carries jobs through to completion. 93.2%. Nod’s score on LongMemEval, a public long-term-memory benchmark — among the best published systems. Apache 2.0. The license under which the code is published on GitHub. Why it matters: this is one of the first personal agents to draw an explicit trust boundary — your machine, your model provider, and a permission check before every irreversible action.

What Nod is — and is not

Nod is a standalone app, separate from n8n and from the agents you build inside n8n. No n8n account is required, no server or cloud service to set up: you install it, connect your accounts, and ask. Its architectural signature is a single sentence: Nod has its own computer and its own browser, in the form of a sandbox — a private virtual machine running locally.

That detail changes what the agent can do. Where a chatbot hands you back a to-do list to execute yourself, Nod can finish the job in the connected apps — Gmail, Calendar, Drive, Slack — and on websites that offer no integration at all. The examples the vendor publishes are concrete: cancel a subscription while declining the retention offers, watch a price and alert you when it drops, download invoices from a supplier portal each month and file them, or pull the numbers from a dashboard that exports nothing.

Nod also works ahead of you: it reads the connected apps, offers the jobs it could take off your plate, and can run several of them in parallel in the background while you talk to it in the app or over Telegram and WhatsApp. Recurring jobs — a morning brief, a price check — can run on a schedule.

The trust boundary: five explicit guarantees

An agent that can send an email or pay with your card has to earn your trust. Nod structures it around five mechanisms, and that is where it differs from cloud agents.

First, openness: everything is under Apache 2.0, and every rule the agent follows lives in the code. Second, isolation: data, memory, and saved logins stay on the machine running Nod, and the model provider only receives what each task requires. Third, permission: before sending, paying, publishing, or deleting anything you did not ask for, a separate check confirms the order came from you — a web page or an email cannot give Nod orders. Fourth, opaque secrets: saved logins and cards are encrypted on the machine, and Nod enters them through private forms the model cannot read. Fifth, traceability: every sign-in, card use, and action is logged, along with the cost of each task.

The most honest part of n8n Labs’ write-up concerns the limit of that control: the permission model is itself a model, and therefore fallible. That is exactly why Nod asks again when your words do not cover an action. Trust is not announced; it is designed to be verified.

The trade-off is real and worth naming. An agent that can spend money and send mail on your behalf turns your own workstation into the crown jewel. The moment Nod holds your credentials, your machine — not a vendor’s cloud — becomes the target worth compromising, and an open-source codebase is simultaneously the best audit trail you could ask for and the clearest map for an attacker. The sandbox and the permission check are the load-bearing walls: they mean a compromised agent still has to ask before it acts irreversibly. That is a weaker guarantee than “nothing can go wrong”, and a much stronger one than “trust us, it is in the cloud”.

The model provider is a choice, not a dependency, and that inverts the usual lock-in. You can switch from one API to another without migrating the agent, or point Nod at an open-weight model running on your own hardware and keep everything on-premises. For a freelancer who would rather not ship client data through a chat service, that inversion is not a feature on a list — it is the product.

The economics: open source, with a bring-your-own model provider

Nod is free and open source. The real cost is the AI you consume: you connect your own provider — an API key, your ChatGPT plan, or an open-source model running locally on your machine. n8n users can also connect their instance so Nod can lean on their workflows.

The vendor concedes that some early testers found tasks more expensive than expected, and that cost reduction has been a focus of recent releases. The guardrails are built to stop the bill from running away: every task shows what it cost and ends with a receipt; you can set a monthly spending limit; Nod checks in once a task passes $5 (a threshold you can change); and each scheduled task displays its estimated monthly cost.

Memory is the other pillar. Nod remembers months of conversation, learns your preferences as it goes, and, once connected to Gmail, mirrors how you write so its drafts sound like you. It also records how it accomplished a task and reuses those notes next time, fixing them when a site changes. All of it remains visible, editable, and deletable by you.

What Nod does not do yet

The research preview is frank about its limits. Nod stops when the computer sleeps: some testers keep it on an always-on Mac mini or on their own server, and a hosted version is on the roadmap. Each installation has a single owner — no multi-user use. And Nod cannot make phone calls.

Those limits draw the product’s exact perimeter. Nod is not a team agent or a contact center: it is a solo operator — a freelancer, an independent, or a household delegating paperwork, follow-ups, and repetitive logins. None of these gaps are defects so much as a clear statement of scope: Nod is being shaped as a single-player, single-machine tool first, and the roadmap’s hosted version is where the multi-user story would begin.

The verdict

If you are a freelancer or solo operator whose weeks vanish into forms, follow-ups, and supplier portals, try Nod on an always-on machine — the “your machine, your model provider, explicit permission” model is the most defensible formulation of the personal agent today, and it carries no subscription. Start with reversible tasks and cap the monthly spend: per-task cost is the one variable you do not yet control.

If you handle regulated data or corporate secrets, stay cautious: local isolation is a real plus, but a research preview whose permission check is itself a model does not go onto a production machine without review. And if you want a multi-user team agent or a hosted service, wait for the announced hosted version — as it stands, Nod is a single-player promise, not a platform.

References

The cyber brief, every Tuesday

The flaws that matter and the patches to apply, in a ten-minute read.

No spam. One-click unsubscribe.
read next

On the same topic

← Back to the feed

Type at least two characters.

↑ ↓ navigate ↵ open esc dismiss