Automation & AI · · That Guy Sketch

n8n Nod: A First Look at the Open-Source Personal AI Agent You Can Run Yourself

n8n Labs just released Nod, an open-source personal AI agent for Mac, Windows and your own server. What it does, which models it runs, and why some local models fail.

Nod is a free, open-source personal AI agent from n8n Labs that runs on your own computer or server, and you choose the model behind it: a cloud model, your ChatGPT plan, or a model running locally. It is a research preview, so it has rough edges. We have been using it for calendar management, and it is already useful. Here is what it does, what worked for us, and where it broke.

Facts below are as of November 16, 2026, from n8n’s launch post, the official Nod docs and third-party coverage. Nod was released on November 10, 2026, and it is changing fast, so check the official docs before you rely on any detail.

What Nod is

Nod is a standalone app, separate from the n8n workflow tool, though you can optionally connect an n8n instance. Per n8n’s announcement, it is released under the Apache 2.0 license and runs on Mac, Windows, or a server you host. The source is on GitHub.

An AI agent is software that plans steps, uses tools such as email, a calendar and a browser, and carries out a task for you. It does more than answer questions. Nod works in the same space as agents we covered in our look at Hermes, Muse and OpenClaw.

What n8n says it can do:

  • Connect to your apps. Gmail, Calendar, Drive and Slack, plus any website, including sites with no integration, through its own browser.
  • Take requests where you are. In the app, on Telegram or on WhatsApp.
  • Work in a sandbox. It runs in a private local virtual machine with its own browser, not loose on your desktop.
  • Ask before it acts. A separate check confirms you asked for it before it sends, pays, posts or deletes anything, and every action is logged.
  • Show you the cost. Tasks display what they cost, you can set a monthly spending limit, and Nod checks in when a task passes $5.

Which models can run it

This is the part we like most. Nod is “bring your own model,” and per the official docs you can connect:

Option How it connects What to know
Claude (Anthropic) API key An API connection is required. A Claude subscription sign-in is not an option.
OpenAI API API key Choose Global or Europe region.
ChatGPT plan Sign in with ChatGPT Needs Plus or Pro. Usage comes from your plan’s included allowance, not API billing.
OpenRouter Sign-in or API key Only models that support text and tools are listed.
Local models OpenAI-compatible server Ollama, LM Studio, vLLM, llama.cpp and similar.

Signing in with a ChatGPT plan is the neat one. If you already pay for ChatGPT Plus or Pro, you can point an agent at it and use the credits you already have, instead of setting up separate API billing.

What we are using it for

Calendar management. Checking what is coming up, finding time, and moving things around, with Nod asking before it changes anything. It is a small, bounded job, which is the right kind of task to hand to a preview-stage agent first.

Our view: this is a nice self-hosted option. You run it where you choose, you pick the model, and the code is open.

The downside: not every local model works

This is where it got messy for us, and it is worth knowing before you spend an evening on it.

What we saw. GPT-OSS 20B loaded and worked fine. Gemma 4 failed when we tried to load it.

Why this probably happens. We have not confirmed the exact cause for Gemma 4, so treat this as informed guessing. The official local-models page lays out two hard requirements that many models and default setups miss:

  1. Tool-call support. Nod does everything through tools. A model without tool support can only chat. The docs say to look for models marked “Tools” on ollama.com. GPT-OSS 20B is the one example model the docs use, which fits what we saw.
  2. A big context window. Nod needs at least 64,000 tokens of context. Its own instructions and tools use about 20,000 tokens before your first message, and many servers default to 4,096 and silently cut those instructions off. The result can look like ignored requests, lost tools or repetition.

If a model fails to load or behaves strangely, check these first:

  • Does the model support tool calls on your server and in your model version or quantization?
  • Is the context length set to 65,536 or more, and has the model been reloaded since?
  • Can Nod reach the server? Test with curl http://localhost:11434/v1/models for Ollama.
  • If Nod runs in Docker or on another machine, remember that localhost means Nod’s own host, not yours.

Server settings from the docs:

Server Context setting Tool calls
Ollama OLLAMA_CONTEXT_LENGTH=65536 ollama serve Filter for models marked Tools
LM Studio Load with context length 65,536 or more Use a tool-capable model
vLLM --max-model-len 65536 --enable-auto-tool-choice --tool-call-parser <parser>
llama.cpp -c 65536 --jinja

The docs do not publish a list of tested or recommended models, so expect some trial and error. If you hit a model that fails, tell us which one and what the error was.

Other limits to know about

  • It is a research preview. n8n says so, and reviewers describe rough edges.
  • It stops when your laptop sleeps. A server install avoids that.
  • Each install has one owner. It is not a shared team tool yet.
  • Local models cost nothing per use, but are slower and less capable. The docs note slow first replies are normal.
  • Safety checks on small models are less reliable. The docs say the approval checks, which use a separate small model or decision service, are less reliable at small sizes.
  • Images only go to vision models on Ollama and LM Studio. Other servers get text only.

Is it fully private?

Be careful here. n8n’s announcement says your data, memory and saved logins stay on the machine running Nod. Some third-party coverage reports that app connections such as Gmail pass through n8n’s hosted connection service and Composio by default. The official docs list “Composio apps” as a connection type, but we could not find a page that explains how Gmail or Calendar are wired. Until that is clear, assume app connections involve a third party unless you verify otherwise, even when the model runs locally.

Who should try it

  • Good fit: people who already like self-hosting, want an agent they can inspect, and have a low-risk task to test, such as calendar work.
  • Wait: anyone who needs it to be dependable for customer-facing or money-moving work, or who needs a team-shared tool.
  • Not the right tool: a business process that must run the same way every time. A fixed workflow in n8n, with a model doing only the bounded steps, is more reliable. See 11 tips for a production agentic workflow.

A sensible way to start: install the desktop app, connect a cloud model first so you know the app works, then try a local model with the settings above.

FAQ

What is n8n Nod?

Nod is an open-source personal AI agent from n8n Labs. It runs on your Mac, Windows PC or your own server, connects to apps like Gmail and Calendar, and asks for approval before sensitive actions.

Is Nod free?

The app is free and open source under Apache 2.0. You pay your model provider directly, unless you use a local model or your own ChatGPT plan allowance.

Can Nod use Claude?

Yes, with an Anthropic API key. Per the docs, a Claude subscription sign-in is not supported for Nod.

Can I use my ChatGPT subscription with Nod?

Yes, with ChatGPT Plus or Pro. Nod uses your plan’s included usage rather than API billing.

Which local models work with Nod?

Models that support tool calls and run with at least 64,000 tokens of context. We got GPT-OSS 20B working and Gemma 4 failing to load. The docs do not list tested models.

Is Nod the same as n8n?

No. Nod is a separate app from n8n Labs. You can optionally connect an n8n instance to it.

Can I rely on it for business-critical work?

Not yet. It is a research preview. Use it on low-risk tasks and keep approvals on.

Want help with a local AI or automation setup?

That Guy Sketch builds automation and AI systems for small businesses, from local models to approval-gated workflows with n8n. Get Started and we will help you pick a setup that fits.

Sources

Read on 2026-11-16. Verify against official pages before relying on any detail.

Something here sound familiar?

Fifteen minutes, no deck, no pitch. Tell us what's annoying you and we'll tell you what we'd do about it.