n8n Nod: Free Open-Source Personal Agent With Its Own Computer

n8n Labs released Nod on November 10, 2026, and the pitch is blunt: a personal agent that does the errand instead of writing instructions for it. Nod is a free, open-source desktop app that runs on your own computer, inside its own private virtual machine, with its own browser. You hand it a job in plain language, it works through the tabs, logins and forms, and it asks before it sends, pays, posts or deletes anything you did not ask for. The app costs nothing. The catch is that every task burns your own model provider, and the bill for that is yours.

That trade is the whole story. Nod gives you local control and your choice of model, and in return you pay the inference cost and keep a machine awake. If you already pay for ChatGPT Plus or run a local model, the deal looks strong. If you expected a free agent that runs itself, it is not that. This review works through what Nod does, what it costs in practice, how its safety checks hold up, and where it falls short.

What Nod is, and what it is not

Nod comes from n8n Labs, a new team inside n8n that works on ideas outside the core product. It is a standalone desktop app, and the first thing to clear up is that it is not n8n. You do not build workflows with nodes, you do not need an n8n account, and there is no server or cloud service to set up first. You chat with it, and it does the work.

It is also not the n8n AI Assistant, and it is not a chatbot with a browser plugin. Nod runs in a private Linux workspace on your machine, with its own Chromium browser, files and terminal, so it can finish jobs on ordinary websites that have no API. That is the difference that matters: a chat assistant hands you a to-do list, while Nod is built to work through the list itself.

The code is published under the Apache 2.0 license at github.com/n8n-io/nod, and the repo had about 1,085 stars four days after launch. It is a research preview, which n8n states plainly on the product page and in the repo. Expect rough edges and frequent updates: the launch announcement in the n8n community forum warns exactly that.

How it works: its own computer, your model

Under the hood, Nod is one Node process and a few Docker containers, all on the machine that runs it. A sandboxed Linux workspace and a Chromium browser container do the actual work, with PostgreSQL storing state. The stack is TypeScript on Node.js 24, built on the Pi agent runtime, with Composio powering app connections and MCP support for extra tools.

You talk to Nod in the app, or through Telegram and WhatsApp, and one conversation carries across all of them. It connects to Gmail, Calendar, Slack, Notion, Drive and Linear, and the app catalog runs to more than 200 apps through Composio. Several jobs can run in parallel, and recurring work such as a morning brief or a price check can run on a schedule.

Nod does not run models itself. You bring your own: an API key for any of more than 30 providers, including OpenAI, Anthropic, Google, OpenRouter, Amazon Bedrock, Google Vertex AI and Azure OpenAI, or a ChatGPT Plus or Pro plan, or a local model through Ollama, LM Studio, vLLM or llama.cpp. A local model needs tool calling and a 64,000-token context window, and n8n warns that small models are less reliable at the safety check, which is the part of the system you least want to weaken.

Setup has five stages, which n8n estimates at about ten minutes: install the app, connect an AI provider, connect your apps with a single sign-in, save the website logins Nod may use, and start asking for things from the app or Telegram. The desktop app needs macOS 14 or later, on Apple silicon or Intel. Windows 11 support is a preview that needs WSL 2, Linux x64 ships as a Flatpak preview, and you can host Nod on your own server with Docker Compose for one person.

What it can actually do

Nod is aimed at people who run their own work: client calls, follow-ups, invoices and supplier portals, plus life admin such as a refund or a booking. The jobs n8n lists on the product page are concrete, and they are the right test of whether this kind of agent earns its keep. On finance, it chases an unpaid invoice until it is paid, with friendly reminders, and asks before sending a second nudge. It pulls supplier invoices each month and downloads bank statements. On life admin, it cancels a subscription, turns down the retention offer, and asks for a refund.

On work, it reads your threads and calendar each morning and writes a one-page brief for each call, it finds the client question you never answered and drafts the reply from your own pricing sheet, and it turns client requests from email and Slack into Linear tasks while skipping duplicates. On monitoring, it checks a page every few hours and tells you only when something changes, for example a client raising prices from $1,200 to $1,500. Two details matter here. The call brief is built from your own messages, calendar and the client's latest news, and stored on the machine that runs it. The draft reply comes from your pricing sheet, not from a guess. That grounding is what a personal agent needs, because a wrong price in a client email is worse than no reply at all.

Nod also keeps months of conversation and records how it completed tasks, reusing those notes and revising them when a site changes. Its memory learns from your own messages only, never from the pages or emails it reads, and every note is visible: you can see, edit or forget each one, with the message it came from. It learns your writing style from sent emails, without names or numbers, so its drafts sound like you rather than like a template.

What it costs: free app, paid model

Nod itself is free, and there is no subscription, no per-seat fee and no credit pack to buy. The money question is entirely about the model you connect, and there are three ways to pay for that.

ChatGPT Plus at $20 a month works out to about £15.14 or CA$28.54, and Pro at $200 a month to about £151.38 or CA$285.42, at November 14, 2026 market rates of 1 USD = 0.7569 GBP and 1.4271 CAD, verified at publication. Those are approximations: OpenAI bills in dollars and your bank sets the final figure. Pro is billed monthly only, with no annual discount, and OpenAI no longer publishes an exact usage multiple for the Pro tiers, so treat the allowance as headroom rather than a number you can plan around.

The API route has no fixed monthly price at all, which is why Nod builds cost controls into the app. Every finished task ends with a receipt showing what it cost. You can set a monthly spending limit, and Nod checks in with you once a task passes $5, about £3.78 or CA$7.14, a threshold you can adjust. Each schedule shows its estimated monthly cost before you turn it on, and the Usage screen records app actions, sign-ins and model costs.

Those controls exist because the cost can surprise you. n8n says some early testers found tasks cost more than they expected, and that the latest release brought costs down substantially, without publishing before-and-after figures. A useful way to budget is to run one representative errand, read its receipt, and multiply by how often you will run it: a $0.40 receipt run done daily is $12 a month before retries, and one retry in four lifts it to about $15. That is arithmetic, not a promise, and your own receipts are the only honest baseline.

The safety model: approval, vault, and a check that can be wrong

An agent that can log in to your accounts and act on websites needs more than a permission toggle, and Nod's design takes that seriously. Before Nod sends, pays, posts or deletes anything outside the scope of your request, a separate server-side check confirms you asked for it. Saved logins and cards are encrypted on your machine, and the vault enters them on websites privately, so the model never sees your passwords. Results from apps, the web and the browser reach the model marked as data, not instructions, which is the defense against prompt injection in a page or email telling the agent to do something else. Every action, sign-in and model cost is logged in Usage, so you can see exactly what Nod did. Every rule it follows is in the open-source code.

There are two honest limits to that design. The safety check is itself a model's judgement, and n8n says so on the product page: it can make mistakes, which is why Nod asks before actions your request does not cover. And your data does not stay entirely on your machine. App connections use n8n and Composio by default, your model provider receives the information each task needs, and n8n receives usage data to improve Nod unless you turn off Share usage data in Settings, under Privacy and data. The product page and the privacy docs list what is sent, and turning the toggle off stops new sends.

The benchmarks, read carefully

Nod's product page leads with four numbers, and they look strong until you read the fine print underneath them. All four come from n8n's own testing, run in October 2026 with GPT-6.1 Sol, in simulated accounts and websites. The README adds the line that matters: test results, not a guarantee.

The 98% figure is the one to be careful with. A synthetic benchmark is a controlled test on simulated accounts, not proof that Nod will finish 98% of jobs on live websites, with real logins and real payment flows. Sites change, logins fail, and a refund flow that works in a test app can stall on a real one. Treat the number as evidence that the agent can complete multi-step errands, not as a success rate you should expect on day one.

The safety numbers deserve the same care, and here n8n published a result that cuts against its own marketing. In 240 test decisions about whether an action was safe, GPT-6.1 Sol and Claude Opus 5.5 got every one right, while Claude Haiku 5.5 let 8 unsafe actions through. That is a useful finding, and it is also a warning: the safety check is only as good as the model running it, and the cheapest model in the list is the one that failed. If you connect a small local model to save money, you are weakening the exact layer that is supposed to protect you. Pick a capable model for the safety check, even if a cheaper one does the routine work.

Nod vs Grok Bot 1.0: free and local vs $30 and hosted

The closest comparison is xAI's Grok Bot 1.0, which launched the same week at $30 a month and joins Slack and Microsoft Teams as a teammate. The two products answer the same question, who does the errand, in opposite ways.

The choice comes down to what you value more. Nod wins on privacy, cost control and model choice, and it is the only one of the two you can inspect or self-host. Grok Bot wins on convenience: it is always on, it lives where your team already works, and there is no machine to keep awake and no model bill to watch. If you want an agent for personal errands and you already pay for a model, Nod is the cheaper and more private option. If you want a teammate for a whole Slack workspace, Nod is not that product yet. Our Grok Bot 1.0 review covers the hosted side in detail. The same trade applies against cloud agents in general. OpenAI's ChatGPT Wallet, which we covered this week, lets agents pay without your card number, but it still runs on OpenAI's infrastructure with OpenAI's model. Nod is the local-first answer to that pattern: your machine, your model, your rules, and your bill.

How to set it up, and the limits that matter

Download the app from nod.n8n.io, connect an AI provider, connect your apps with a single sign-in, save the website logins Nod may use, and ask for something in the app or on Telegram. Set the monthly spending limit before your first real errand, not after the first surprising receipt. If you run a local model, check that it supports tool calling and a 64,000-token context window, and remember that small models are less reliable at the safety check. The official docs at nod.n8n.io/docs cover setup and common problems, and the n8n community forum is where early users share what works.

Nod is a research preview, and its limits are structural rather than cosmetic. It stops when the laptop running it sleeps, so nothing runs overnight unless you put it on an always-on Mac mini or your own server. Each installation has one owner, so it cannot be the shared automation for a team. It cannot make phone calls, which rules out the errands that still need a voice. A managed hosted version, phone calls, team use and custom skills are all on the roadmap, not available now. The Windows and Linux builds are previews, so Mac is the safest bet for a first install. And the benchmark figures are n8n's own, run on synthetic accounts with one model, with no independent replication yet. None of that makes Nod a bad product. It makes it a preview you should try with low-stakes errands first, and only hand it your money and your inbox once you have seen it finish a few jobs cleanly.

Three mistakes are easy to make. Connecting the cheapest model and expecting the safety check to hold, when n8n's own test showed a small model letting unsafe actions through. Skipping the spending limit, because the app is free and it is easy to forget that every task has a meter running. And handing it a high-stakes job first, instead of starting with a price watch or a call brief, reading the receipt and the Usage log, and only then giving it an invoice chase or a refund.

Is Nod really free?

Yes, the app is free and open source under Apache 2.0, with no subscription or per-seat fee. You pay your model provider for the AI it uses, either per token on an API key, through a ChatGPT Plus or Pro plan, or nothing per token if you run a local model on your own hardware.

Do I need n8n to use Nod?

No. Nod is its own app from n8n Labs, and it does not require an n8n account or a server. It is not the n8n AI Assistant and it does not build n8n workflows, although it can draft workflows in n8n if you connect it, and they stay drafts until you ask it to publish.

Which computers can run Nod?

Mac with macOS 14 or later, on Apple silicon or Intel, is the full release. Windows 11 is a preview that needs WSL 2, Linux x64 ships as a Flatpak preview, and you can host Nod on your own server with Docker Compose for one person. Nod needs the computer running it to stay on and awake.

Is it safe to give Nod my logins?

Saved logins and cards are encrypted on your machine, and the vault enters them on websites without showing them to the model. A separate check confirms you asked for anything Nod sends, pays, posts or deletes. Even so, the safety check is a model's judgement and can be wrong, so start with low-stakes jobs and read the Usage log after each one.

What can Nod not do yet?

It cannot make phone calls, it stops when your laptop sleeps, and each installation has one owner, so there is no team use yet. A managed hosted version, phone calls and a team version are on n8n's roadmap. It is a research preview, so expect rough edges and frequent updates.

Nod is the most interesting personal agent launch of the week because it refuses the usual deal. Instead of a hosted agent that rents you intelligence by the month, it gives you the agent for free and lets you bring your own model, on your own machine, under rules you can read in the code. The price is real: you pay the model bill, you keep a machine awake, and you accept a preview. For people who already pay for a model and care where their data lives, that is a fair trade, and the most honest way to find out is to install it and hand it one small errand.

Download it from nod.n8n.io and read the source at github.com/n8n-io/nod before you connect an account. If you try it, start with a price watch, set the spending limit first, and check the receipt.

Sources

• Nod, official product page, nod.n8n.io — release status, features, pricing, benchmarks, system requirements
• n8n-io/nod on GitHub — Apache 2.0 license, architecture, benchmark table, setup, star count
• Nod docs, How we measured this — benchmark methodology, GPT-6.1 Sol, October 2026, simulated accounts
• Nod docs, How trust works — safety check, vault, usage data
• Nod docs, Spending — per-task receipts, monthly limit, $5 check-in
• Unite.AI, n8n Releases Open-Source Personal Agent Nod for Mac and Windows — November 10, 2026 launch, providers, cost controls, limits
• AI Reiter, n8n Nod Review: Local AI Agent Setup, Sandbox and Cost — setup stages, benchmark snapshot, cost arithmetic
• The n8n Blog, Nod by n8n Labs — release summary, providers, safety model
• OpenAI, ChatGPT pricing — Plus $20 a month, Pro $100, $200 and $500 a month