n8n Nod Setup: Model Costs, Local Privacy and Scheduled Tasks

n8n Nod is a standalone personal agent that runs on your computer and uses a model service you choose. It can work across connected apps and websites, but setting it up properly means checking three things first: what leaves your machine, what a task costs, and whether the computer will stay awake.

n8n released Nod as a research preview on November 10. It is separate from the workflow builder, needs no n8n account, and is open source under Apache 2.0. That makes it accessible to an individual who wants help with recurring work without first building a workflow. It does not make a personal laptop a dependable always-on service.

Start with one bounded job

For a first run, choose work where a mistake is easy to spot and cheap to undo. Comparing public supplier pages is a better starting point than giving an unfamiliar agent responsibility for payments or client communication.

Here is an example task specification, not a tested result: “Compare these three public pricing pages. Put the monthly price, included usage and overage rules in a table with source links. Stop after the comparison. Do not create accounts, contact anyone or purchase anything.”

This gives you something concrete to inspect: whether the prices match the pages, whether the agent distinguishes monthly from annual billing, and whether it stops at the boundary. Connecting more accounts before checking those basics increases the possible consequences without improving your ability to judge the output.

Install locally or connect to a server

The desktop installation guide supports macOS 14 or later on Intel and Apple silicon, and Windows 11 x64. A local Windows installation also needs WSL 2, hardware virtualization and at least 20 GB of free disk space. Connecting the desktop client to an existing Nod server does not require that local Windows runtime.

Download the matching installer from the official guide. During onboarding, choose Get started for a local assistant or Connect to a server for an existing installation. Email connection and update emails are optional; connect only what the first task requires.

For local use, closing the window leaves Nod in the menu bar or system tray. Quitting it stops the assistant. A sleeping computer also stops its work. If a scheduled job fails to run, check the host before changing the prompt. A remote server can keep working after the desktop client closes, but the server itself must remain online.

The useful distinction is where the assistant runs. A phone or desktop interface does not remove that host's uptime requirement.

Free software still needs an inference budget

Nod is free, but a paid model provider charges for usage. The launch also supports connecting a locally running open-source model. Nod shows task costs and receipts, supports a monthly spending limit, and checks in when a task passes $5 by default. That threshold can be changed.

Treat the initial limit as a budget decision rather than a sensible default for every job. Five dollars is very different for an occasional research task and a recurring check that runs many times a day. Start with a single run, review its cost and only then decide whether to schedule it.

Model access also depends on the connection method. The model configuration documentation is the place to check the supported service and credentials rather than assuming an existing subscription to any AI app includes API usage.

Local storage does not mean every request stays local

Nod's privacy documentation separates storage from external processing. Model services receive the conversation and material needed for the task, which can include screenshots. Connected apps receive actions, and managed app connections involve Composio and a connection service. A self-hosted deployment also gives the server administrator access to its stored data.

Usage analytics are on by default for new owners; conversation traces are off. Both controls are under Settings → Privacy & data. Turning sharing off stops future sends but does not retrieve data already sent. The documentation also warns that trace scrubbing cannot identify every personal detail.

For sensitive work, inspect the whole path: the host, model provider, app connection and optional diagnostics. Installing an agent locally answers only the first of those questions. Do not treat “local” as a blanket confidentiality guarantee.

Decide what deserves unattended execution

n8n says Nod checks whether actions are authorized, but also acknowledges that the checking model can make mistakes. That is a reason to keep early tasks narrow and review outputs, not a reason to assume every confirmation creates a reliable security boundary.

For a recurring job, define the source, expected output, spending ceiling and stopping condition before enabling a schedule. Keep consequential actions separate from research where possible. A weekly comparison can be useful without also granting permission to buy the cheapest option.

The important result of a first Nod task is evidence that you can supervise it: a correct output, understandable cost and clear record of what happened. Only then does moving it onto an always-on machine solve a problem worth solving.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain agents, developer workflows and product tradeoffs in writing that people can use.

Book a 30-minute call.