Nod is an open-source AI agent from n8n that gets its own computer and its own browser and uses them to get errands done on real websites and in your apps: chasing invoices, preparing calls, cancelling trials. It runs on your Mac, Windows PC, Linux machine or your own server, with the model you choose, and it asks before doing anything your request does not cover. Nod is free under the Apache 2.0 license; you only pay your model provider. n8n introduced Nod on November 10, 2026 as a research preview. Two things are worth knowing up front: Claude only works in Nod with an API key, not with a Claude subscription, and schedules only run while the computer running Nod is on. For always-on use, Nod belongs on a server or a Mac mini. All details as of November 15, 2026.
I work with fifteen AI employees, built my own Linux server with Claude Code and Slack for five of them, and ran n8n in my own Kubernetes cluster for a long time. I installed the Nod server from source on my Mac on November 15, 2026, version 0.7.6. Everything else in this article is based on Nod's repository, docs and website, retrieved on November 15, 2026. Everything I have written about AI agents is collected on my AI agents page.
What is Nod?
n8n describes Nod as an “open source personal agent with its own computer”. It was built by n8n labs, which n8n calls a small team inside n8n that tries ideas outside the core product. n8n adds that most of these ideas get shelved, that Nod is out as a research preview, and that it does not change where n8n is headed.
On November 15, 2026 the repository had 1,479 stars on GitHub. According to n8n, setup takes about ten minutes: install the app, connect a model and your apps, then ask for something in the app or on Telegram.
How Nod works
The core of Nod is a computer of the agent's own. According to the architecture docs, Nod is one Node process plus a few Docker containers on the machine running your assistant. In the desktop app this runs in a private virtual machine on your computer, in a source install in Docker containers, and on a server, there.
Five ideas hold it together:
- Approvals live in the server, not the prompt. A separate check decides which app actions, browser steps and sign-ins may run. It sees your words and the proposed action, never the pages or emails Nod read. Reading and drafting run directly. Sending an email, adding an event or deleting files runs if it matches your request; otherwise Nod asks. Every payment waits for your approval with your card's security code, which Nod never saves.
- Vault for secrets. The server enters passwords, one-time codes and cards into the website itself. The model only sees account names and status, never the secret.
- Memory you can see. Memory notes come only from your own messages, never from pages or emails Nod reads. You can view, edit or forget every note. If you connect Gmail, Nod learns your writing style from your sent emails, without names or numbers.
- Schedules and watching. You can ask Nod to check a page every few hours and speak up only when something changes. Nod checks Gmail for new mail while it is running; Calendar and Slack do not wake Nod on their own.
- Everything on record. Usage lists every app action, every sign-in and, for API models, the costs.
An example from the docs: every morning Nod reads your messages, your calendar and the client's latest news and writes a one-page brief for each call. If it finds an unanswered pricing question, it drafts the reply from your own pricing sheet and sends it only after you say yes.
What n8n's own tests show
n8n publishes its own results, all with GPT-6.1 Sol in October 2026 in simulated accounts and websites. In n8n's own tests, Nod finished 98 percent of 226 errands, each run three times, which makes 678 runs. According to n8n, Nod took no action without permission in any of those runs. For recall of earlier chats, Nod scores 93.2 percent on LongMemEval-S according to n8n, and none of 216 attempts to plant false memories through crafted emails succeeded. n8n itself adds: “Test results, not a guarantee.”
Models and costs
Nod does not tie you to a model. According to n8n, more than 30 providers work, including OpenAI, Anthropic, Google, OpenRouter, Amazon Bedrock, Google Vertex AI and Azure OpenAI, plus your own server with an OpenAI-compatible interface. Local models run through Ollama, LM Studio, vLLM or llama.cpp; they need tool calling and a 64,000-token context window, and according to n8n small models are less reliable at the safety check.
Claude via API only. A Claude subscription does not work in Nod. The docs say: “Anthropic's authentication conditions also require an API connection for Claude in Nod.” So you need an Anthropic API key and pay per use. With a ChatGPT subscription, on the other hand, you sign in directly. According to the docs, a Gemini app subscription is not a sign-in option either.
Your model choice is a safety question. The approval check is itself a model. In 240 test decisions, according to n8n, GPT-6.1 Sol and Claude Opus 5.5 got every one right, while Claude Haiku 5.5 let 8 unsafe actions through. Background work such as memory upkeep uses a smaller model from the same provider according to the docs, for example claude-haiku-5-5 or gpt-6-luna.
Costs. Nod itself is free. Your model provider bills you directly. With an API model, Nod shows what each task cost, and at your monthly limit it stops schedules, tasks and background work before calling the model; according to the docs, work that's already running can finish. According to the docs, the small approval checker can also cause API costs. For an OpenAI project with EU data residency, you can choose the region “Europe (EEA + Switzerland)” in Nod; requests then go to eu.api.openai.com. What each model option costs per month, with worked examples, is in Nod Pricing: What the AI Agent by n8n Costs.
Three ways to run Nod
The desktop app needs neither Node.js nor Docker. If you close the window, Nod keeps running in the menu bar; if you quit the app or the computer sleeps, nothing runs. The step-by-step install on a Mac, on Windows and from source is in Install Nod: Set Up the AI Agent by n8n.
According to the docs, one server belongs to one person: whoever has the access token has full access to the assistant and all connected accounts. Prebuilt server bundles sit in private releases according to the docs; the public route is building from source with Docker Compose. Telegram and the WhatsApp QR connection work without opening a port on your router, according to the docs. How to set Nod up on your own server or a Mac mini is in Nod Self-Hosted: AI Agent by n8n on Mac mini.
Nod, n8n workflow or n8n agent
The name is easy to misread, because technically Nod has nothing to do with n8n workflows. n8n says so itself: Nod is its own app, does not need n8n and is not the n8n AI Assistant.
I summarise the first two columns from my articles on n8n and the third from the Nod docs. How to build a workflow is in n8n Workflow Tutorial: Build, Test, Export, how an agent is built in n8n in n8n AI Agent: What I Learned Building One. If you need a run that goes the same way every day, a workflow is usually the more robust choice. Nod plays to its strengths on errands that look a little different each time and lead through logins and forms.
Limits n8n names itself
- Research preview. According to n8n, Nod is an experiment. The approval check is a model's judgement and can be wrong. That is why Nod asks before anything your request does not cover and keeps a record of what it did.
- Schedules only while the computer runs. If the laptop is closed or asleep, nothing runs. For always-on use, n8n points to your own server or a Mac mini that stays on; a hosted version is on the list for later.
- Connection service from n8n and Composio. App connections go through a service hosted by n8n and through Composio by default. For covered calls, Composio has Zero Data Retention switched on: content does not end up in its execution logs, but tool name, time and status do.
- Usage data on by default. Nod sends usage data to n8n, such as features used, timings, costs and failures, not your messages. You can turn it off in Settings → Privacy & data. Nod shares conversation traces only if you switch them on.
- One server, one person. There is no registration or multi-user setup according to the docs. “Nod for your team” is on the list of next steps.
- No separate history deletion. According to the docs, you cannot clear only your chat history yet, only all data at once.
When Nod fits and when it does not
n8n says it built Nod for people who run their own work: client calls, follow-ups, invoices, supplier portals, plus life admin such as a refund or a booking. That leads to a simple assessment:
What sets Nod apart from ChatGPT Dots, Grok Bot and Claude Cowork: Nod is open source under Apache 2.0, and you keep full control. Nod runs on your computer or server, files, memory and logins stay there, and the approval check runs in your own Nod server. You can use any model, from more than 30 providers, through your own server or locally. With Dots, Grok Bot and Cowork, the vendor decides on the model and the platform. The downside: Nod has no flat-rate subscription. Unless you use a ChatGPT Plus or Pro subscription or a local model, you pay per use through the API, and that is the only way to use Claude. The full comparison with ChatGPT Dots, Grok Bot, Claude Cowork, ChatGPT Work and my own AI server is in Claude Cowork Alternative: Nod Compared.
For businesses with customer data: Nod keeps its database, files, memory and logins on the computer or server that runs it. Your model provider, however, receives your conversation, what Nod reads for you, and screenshots. App connections go through n8n and Composio by default, web searches through Exa, or OpenAI with a ChatGPT plan, and usage data goes to n8n. If you send customer data through Nod, check these recipients beforehand, just as you would with any other AI service. What Nod sends to whom and what you can switch off is in Nod Privacy: What the AI Agent by n8n Shares. This is not legal advice. The questions an agent raises when it looks into your inbox on its own are collected in Claude Duet and GDPR: An Agent in Your Inbox; how I separate what an AI may see at all is in AI and Privacy: What the AI Gets to See.
I compare more agents that run on your own machine in Running AI Agents Locally: What Works.
Frequently asked questions
What is Nod by n8n?
An open-source AI agent under the Apache 2.0 license that n8n labs introduced as a research preview on November 10, 2026. Nod has its own Linux workspace and its own browser, connects to your apps and asks before actions your request does not cover.
Is Nod free?
The software is. Costs come from the model provider: pay per use with an API key, or through your ChatGPT Plus or Pro subscription. A local model costs nothing per request.
Can I use Nod with Claude?
Yes, but only with an Anthropic API key. According to the docs, Anthropic's conditions require an API connection; you cannot sign in to Nod with a Claude subscription.
Do I need n8n for Nod?
No. Nod is its own app from the team behind n8n. You need neither an n8n account nor an n8n installation.
Does Nod run when my laptop is closed?
No. Schedules only run while the computer running Nod is on and awake. For always-on use, n8n recommends your own server or a Mac mini that stays on.
Is Nod safe?
Nod checks app actions, browser steps and payments in the server and keeps passwords and keys away from the model. The check itself is a model and can be wrong, and Nod is a research preview. How agents are protected against injected instructions is covered in Prompt Injection: Protecting AI Agents.
What to do next
Before you install Nod, write down one errand you would like to hand off, with a clear end and a clear limit: what should be done and what Nod may not do without you. Then decide which model to use, because costs and the reliability of the approvals depend on it, and whether Nod on your own computer is enough or it should run around the clock. How I build AI employees with clear rights and approvals is something I show in my community Claude Practitioners.
