DIRTY STATIC / DISPATCH Tools18 NOV 2026 • 11 MIN READ
A small sealed glass machine room on a concrete desk in a brutalist hall, with a tiny agent working inside it and caution yellow hazard stripes on the desk edge
ToolsINTELLIGENCE FOR THE MACHINE AGE

Nod: n8n Labs Ships an Open Source Personal Agent

Nod is a standalone desktop app from n8n Labs, released on November 10, 2026 as a research preview. It is a personal agent with its own computer and its own browser, running on your machine, working in Gmail, Calendar, Drive, Slack, Notion and more than 200 other apps, and asking before it acts on your behalf. This profile covers how the sandbox and the trust check work, which models you can bring, what a task costs, what the system needs, and where the limits of a research preview still show.

Start reading ↓

Most chat assistants stop at advice. You ask for help with an invoice, a booking or a form, and you get back a neat list of steps that you still have to do yourself, tab by tab, login by login.

Nod is built to finish that kind of job instead. n8n Labs released it on November 10, 2026 as a research preview. It is a standalone desktop app, free and open source under Apache 2.0, and it runs on your machine with a model you choose. You tell it what you need in plain language, in the app or on Telegram or WhatsApp, and it works in your apps and on real websites until the job is done or it needs your word to go further.

Nod is not n8n. There are no workflows to build, no nodes to wire, and no n8n account is needed. n8n Labs is a new team inside n8n that works on ideas outside the core product, and Nod is its bet on work that belongs to one person rather than to a team sharing one automation system.

A personal agent with its own computer and browser, built for one person and run on that person machine. Primary use case: errands and admin work that spans apps and websites, such as chasing invoices, preparing call briefs, cancelling subscriptions and filing paperwork, with an approval check before it acts for you.

Key features: a private virtual machine sandbox and a separate browser; connections to Gmail, Calendar, Drive, Slack, Notion and more than 200 apps through Composio; persistent memory you can see, edit or delete; scheduled jobs; a Usage view that logs actions, sign ins and model costs; and support for models from OpenAI, Anthropic, OpenRouter and other providers, a ChatGPT Plus or Pro plan, or a local model.

Advantages: the code is open source, data and saved logins stay on the machine that runs it, and the model choice stays yours. Limitations: it is a research preview, it stops when the laptop running it sleeps, each install has one owner, and it cannot make phone calls.

Platforms
macOS 14 or later; Windows 11 in preview; Linux in preview; or a server you host yourself
Pricing
Free and open source under Apache 2.0; model provider fees apply, or use a ChatGPT Plus or Pro plan or a local model at no API cost
Official Nod collage showing the personal agent at work across apps and websites
Nod at work, from the official product site. The agent plans a job, then uses its own computer, its browser and the apps you connect. Image: n8n Labs

What it is

Nod is a personal agent for a single owner. The jobs n8n shows for it are ordinary and stubborn at the same time. Chase an invoice until it is paid, with a check with you before a second reminder. Cancel a subscription, turn down the offers meant to keep you, and ask for a refund. Read your threads, your calendar and a client latest news each morning, then write a brief for each call. Watch a pricing page and speak up only when the price changes. Download statements from supplier portals each month and file them by name.

None of those jobs is hard on its own. Together they eat an afternoon, mostly in logins and forms. A chat tool can draft the words for you. Nod is meant to do the clicking, the filing and the checking as well, with several jobs running in parallel.

How it works

The design choice that sets Nod apart is the second computer. Nod does not take control of your desktop. It works inside a private virtual machine on your machine, a Linux workspace sandbox, with its own Chromium browser beside it. Files, memory and saved logins stay on that machine. Your model provider receives the information a given task needs.

The README describes the stack in plain terms. One Node process and a few Docker containers run on your machine. The server holds the agent runtime, the trust check, the Vault, memory and schedules. Separate containers provide the workspace and the browser. Provider keys never enter the workspace or the browser, which is a sensible line to draw.

Most app connections go through Composio, the integration service behind the catalogue of more than 200 apps, and by default app connections use n8n and Composio. A few services, such as Notion and Linear, connect through their own MCP servers instead. MCP is the open standard many tools now use to expose actions to an agent. The browser covers the rest, signing in with a saved login when a portal has no useful integration at all.

The Vault handles secrets. Saved logins, codes and cards are stored on your machine and entered through private forms the model cannot read, per the n8n announcement. The model can ask for a login to be used. It never sees the password itself.

Official Nod screenshot collage showing a one page call brief prepared from messages, calendar and client news
A morning call brief, from the official product site. Nod builds it from your threads, your calendar and the client latest news, then stores it on the machine that runs it. Image: n8n Labs

Memory, and the number n8n reports for it

Nod remembers. Tell it once who handles invoices now, and it uses that fact from then on. Every note it keeps is visible, with the message it came from, and you can edit a note or tell it to forget one. It learns a writing style from your sent email so drafts sound closer to you, and it writes down the steps it used for a task so it can reuse them and repair them when a site changes. n8n adds that it learns about you from your own messages, never from the pages it reads on your behalf. That boundary is the right one.

For scale, n8n points to LongMemEval, a public benchmark for long term memory in agents. On LongMemEval-S, n8n reports a score of 93.2 percent for Nod, from a test of 500 questions about earlier chats asked in new sessions. Treat that figure exactly as labelled here. It is n8n own reported score, not a result measured by this magazine, and memory in your own inbox will be messier than memory in a benchmark set.

n8n reports two other results with the same care needed. In its own errand tests, 226 errands were each run three times, for 678 runs, with 98 percent finished and no run acting beyond what the user allowed. In an attack test, 108 published attack emails were each tried two ways, for 216 attempts, and none planted a false memory. The README says those runs used GPT-6.1 Sol in October 2026 in simulated accounts, and calls them test results, not a guarantee.

It asks before it acts

An agent that can send email, spend money and delete things needs a brake that is not just a polite prompt. In Nod, the brake lives in the server rather than in the model prompt. The server checks each app action, and each website step that can change something, against your request and your rules. Anything those do not cover waits for your approval. A web page cannot give Nod orders, because results from apps, the web and the browser reach the model marked as data, not as instructions.

n8n admits the limit in the same breath. The safety check is itself a model judgement, and it can make mistakes. In 240 test decisions in the README, two large models got every decision right, while a smaller model let eight unsafe actions through. Pick a capable model is part of the safety story here, and a weak local model weakens more than answer quality.

Cost has a brake too. Every task shows what it cost in Usage, a finished task ends with a receipt, and you can set a monthly spending limit for API models. Nod checks in with you once a task passes $5, and you can change that threshold. Each schedule also shows its estimated monthly cost. n8n says some early testers found tasks cost more than they expected, and that recent releases brought the cost down. No figure is given for that reduction, so read it as a direction, not a saving you can budget around.

What you need to run it

Requirements

Release
Research preview, released November 10, 2026 by n8n Labs
Licence
Apache 2.0, free and open source
macOS
macOS 14 or later, Apple silicon or Intel
Windows
Windows 11 in preview, needs WSL 2
Linux
Linux x64 in preview, or run Nod on your own server for one person
Model
OpenAI, Anthropic, OpenRouter or another supported provider, a ChatGPT Plus or Pro plan, or a local model
Local model needs
Tool calling support and a context window of 64,000 tokens, per the Nod docs
Apps
Gmail, Calendar, Drive, Slack, Notion and more than 200 apps through Composio

Setup, per the product site, is install the app, add an AI provider, connect your apps with one sign in each, and save any logins Nod may use on websites. The docs name more than 30 providers, including OpenAI, Anthropic, Google and OpenRouter, plus local runtimes such as Ollama and LM Studio. A local model must support tool calling and hold a 64,000 token context, and n8n warns that small models are less reliable at the safety check above.

Where your data goes deserves a straight read before you connect a real inbox. Your database, files, memory and saved logins stay on the machine that runs Nod. Model services, apps and websites receive the data they need for your tasks. App connections use n8n and Composio by default. Beyond update checks, n8n receives usage data to improve Nod, and you can turn that off in Settings.

What Works

  • Free and open source under Apache 2.0, with its rules visible in the code
  • Runs in its own sandbox with its own browser, rather than taking over your desktop
  • Asks before actions your request does not cover, and logs actions and costs in Usage
  • Model choice stays wide and yours, from major providers to a local model

What Does Not

  • Research preview, with rough edges and a shape that will change
  • Stops when the laptop running it sleeps
  • One owner per install, no phone calls, and no team version yet
  • App connections run through n8n and Composio by default

Which should you choose

Choose Nod if the errands are yours alone. Solo operators and freelancers whose week is full of follow up work across an inbox, a calendar and a pile of portals are the reader n8n built this for. The value is not a clever answer. It is a finished job, with a receipt and a record you can check.

Choose Nod if you want the agent on your own machine. If the alternative is a cloud agent that holds your logins on someone else server, Nod offers a different trade. Your files and memory stay local, the code is open for inspection, and you can run it on your own server.

Stay with a chat assistant if you only need words. Drafts and summaries do not need a sandbox, a browser or a Vault. Nod earns its setup only when a job has to touch real apps and websites and come back done.

Stay with n8n itself if the work is shared. Team processes other people depend on belong in n8n, where they already live. Nod does not replace that system.

Wait if you need it awake around the clock with no effort. A hosted version is on the roadmap, not in the download. For now the options are a machine that stays on, such as a Mac mini, or a server you host yourself.

Final thoughts

Nod is an interesting release because its claims are mostly about restraint. It has its own computer so it does not need yours. It asks before it acts so a stray page cannot act through it. It shows its costs and keeps its memory visible. For a category that loves a grand promise, that is a quieter and better pitch.

It is still an experiment. A research preview that stops when your laptop sleeps, depends on a model you supply, and leans on a model based safety check will get some jobs wrong in the wild. The question Nod puts to you, with one real job from your own list, is whether the jobs it finishes cleanly are worth the supervision the rest still need. At least it gives you the logs to answer that yourself.

Sources

  • Release, purpose, jobs, memory score, trust design and cost guardrails: Meet Nod, n8n Blog, November 10, 2026
  • Product overview, setup, models, platforms, data handling and limits: Nod, official product site
  • Licence, architecture, test figures, model safety comparison and source code: n8n-io/nod on GitHub
  • Call brief and product collage images: n8n Labs, via the official product site
  • Hero illustration: Dirty Static, generated in house style
■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘