n8n Labs released Nod on 10 November 2026. It is a free, open-source personal AI agent that runs on your own Mac or Windows PC, with its own private computer and browser.
You bring the model. Nod works with Anthropic, OpenAI, Gemini, OpenRouter and many more, or with a local model. This guide covers install, setup, safety, costs and fixes in one place.
- What: a standalone personal agent from n8n Labs, separate from the n8n workflow product. No n8n account needed.
- License: Apache 2.0, code at github.com/n8n-io/nod. It is a research preview.
- Platforms: macOS 14+, Windows 11 x64 (preview, uses WSL 2), Linux x64 Flatpak preview, or self-host on a server.
- Models: API keys from OpenAI, Anthropic, Gemini, OpenRouter, Bedrock, Vertex AI, Azure and 20+ more, a ChatGPT Plus or Pro sign-in, or a local model via Ollama.
- Cost: Nod is free. You pay only your model provider. A $200 monthly limit is set by default and it checks in when a task passes $5.
- Safety: a separate trust check asks before sends, payments, posts or deletes you did not request. Logins and cards sit in an encrypted Vault the model cannot read.
What Nod is (and is not)
n8n calls Nod “a personal assistant with its own computer”. It lives on your machine, does jobs in your apps and on websites, and keeps going until the job is done. A chat tool would only hand you a to-do list.
It comes from n8n Labs, the team that runs experiments outside the core n8n product. n8n says its main workflow product and its own agent work are unchanged.
- Not n8n itself: you do not need an n8n account, server or cloud setup to use Nod.
- Single owner: each install serves one person, not a shared team.
- Optional n8n link: if you connect an n8n instance, Nod can build, test and run workflows. New workflows stay drafts until you tell it to publish.
Note: if you want a classic n8n automation stack with a local model instead, see my n8n self-host with Ollama AI agent guide. Nod is a different, chat-first product.
What Nod can do
Under the hood, Nod is one Node.js process plus a few Docker containers on the machine running it. These include a Linux workspace sandbox, a Chromium browser, PostgreSQL and the Pi agent runtime. Your provider keys stay out of the sandbox and browser.
Built-in skills
- Research: search the web, read pages and PDFs, and compare sources.
- Files: write documents, sheets and code inside its own Linux workspace.
- Apps: read and draft in Gmail, Calendar, Drive, Slack, Notion, Linear and more.
- Browser: fill forms, sign in with saved logins and download files, even on sites with no API.
- Schedules: recurring jobs such as a morning brief or a daily price check.
- Monitoring: when asked, watch for new email that needs your attention.
Where you can talk to it
You can chat in the desktop app, in Telegram or in WhatsApp, all in one conversation. Nod can run several jobs in the background at once and suggest jobs based on your connected apps.
Example jobs n8n lists
- Money: cancel subscriptions and chase refunds.
- Shopping: watch prices on a product.
- Travel: book a hotel after you approve it.
- Admin: download supplier invoices and export data from dashboards with no export button.
Memory is a big selling point. n8n says Nod remembers months of chats and scored 93.2% on LongMemEval, a long-term memory benchmark. You can view, edit or delete anything it stores about you.
System requirements
| Platform | Requirement | Download |
|---|---|---|
| Mac (Apple silicon) | macOS 14 or later, about 1.8 GB for app and runtime | nod-arm64.dmg |
| Mac (Intel) | macOS 14 or later | nod-x64.dmg |
| Windows | Windows 11 x64, preview. The private computer runs as a WSL 2 distribution | nod-Setup-x64.exe |
| Linux | x64 only, Flatpak, GNOME Keyring or KWallet, hardware virtualization with /dev/kvm |
Flatpak parts (preview) |
| From source | macOS or Linux, Node.js 24.21+, pnpm and Docker | GitHub repo |
You also need model access: an API key, a ChatGPT plan that supports sign-in, or a local model server such as Ollama. The desktop app does not need Node.js or Docker installed.
How to install Nod on Mac, Windows and Linux
All builds are on the latest release page of the n8n-io/nod GitHub repo, linked from nod.n8n.io.
Mac
- Check your chip. Open Apple menu, then About This Mac. Look for “Chip” (Apple silicon) or “Processor” (Intel).
- Download the right DMG. Get
nod-arm64.dmgfor Apple silicon ornod-x64.dmgfor Intel. - Install. Open the DMG, drag Nod to Applications, then open it. Click Open when Gatekeeper asks about an app from the internet.
- Allow the keychain. If macOS asks for your keychain password, approve it so Nod can protect saved keys.
Windows
- Download. Get
nod-Setup-x64.exefrom the latest release. - Run it normally. Use a normal PowerShell window, not Run as administrator. The installer checks the installer and runtime, installs Nod and opens it.
- Let WSL set up. During the preview, Nod’s private computer runs as a WSL 2 distribution, so keep WSL enabled.
Linux (Flatpak preview)
- Add Flathub. Run the command below.
- Download the parts. Put every
nod-linux-x64.flatpak.partNNfile andSHA256SUMS-linux-x64in one empty folder. - Join, verify and install. Run the second command in that folder.
- Launch. Open Nod from your app menu or run
flatpak run io.n8n.Nod.
flatpak remote-add --user --if-not-exists flathub https://dl.flathub.org/repo/flathub.flatpakrepo
cat nod-linux-x64.flatpak.part[0-9][0-9] > nod-linux-x64.flatpak && sha256sum --check SHA256SUMS-linux-x64 && flatpak install --user ./nod-linux-x64.flatpak
From source (macOS or Linux)
git clone https://github.com/n8n-io/nod.git
cd nod
corepack enable
pnpm start
First launch
- Pick a mode. Choose Get started for a local assistant, or Connect to a server if you self-host Nod elsewhere.
- Answer setup. Optionally connect email and answer a few questions about you.
- Add a model. Connect a provider if none is set (next section).
- Choose privacy. Basic usage data is shared by default for new owners. Turn it off in Settings, then Privacy & data.
- Wait for the runtime. Nod downloads its runtime with a progress bar, so stay online.
Tip: Nod stops when your laptop sleeps. Some testers run it on an always-on Mac mini or their own server so schedules keep firing.
How to connect your model keys
Nod is “bring your own model”. Requests go straight from Nod to the provider you pick, and that provider bills you.
- Open model settings. Click Settings at the bottom of the sidebar, then Models.
- Add a connection. Click Add connection (or Add), then choose a provider.
- Use an API key. Pick Use an API key. Leave API region on Global unless you need EU data residency.
- Paste and connect. Paste the key into API key and click Connect.
- Pick the model. Click the model name after “When you talk to Nod, it uses” and choose from the list. Recommended models appear first.
ChatGPT can connect with an API key or your existing ChatGPT Plus or Pro sign-in. OpenRouter supports an API key or browser sign-in. Need a key? See how to get a Claude API key or my OpenRouter guide with free models.
Recommended starting models (from n8n’s docs)
| Connection | Default pick |
|---|---|
| Anthropic | Claude Opus 5.5, high effort |
| OpenAI / ChatGPT | GPT-6.1 Sol or GPT-6 Sol, high effort |
| Google Gemini | Gemini 3.8 Flash, medium effort |
| OpenRouter | GPT-6.1 Sol, high effort |
| Google Vertex AI | Gemini 4 Argon, medium effort |
| Amazon Bedrock / Azure OpenAI | Your chosen model or deployment |
You can also set a different model for schedules and background work, such as “The light model”. That keeps recurring jobs cheaper.
Warning: model choice affects safety, because the trust check is also a model. n8n’s own tests found Claude Haiku 5.5 let 8 unsafe actions through in 240 test decisions, so pick a capable model for real tasks.
Run Nod with a local model (Ollama)
Yes, Nod supports local models through any OpenAI-compatible server. The docs give steps for Ollama, plus addresses for vLLM and llama.cpp.
- Tools: the model must support tool calls, because Nod works only through tools.
- Context: use at least 64,000 tokens. Many servers default to 4,096, which silently cuts Nod’s instructions.
- Reachability: the server must be reachable from the machine running Nod.
- Pull a tool-capable model. Run
ollama pull gpt-oss:20b(or filter by Tools on ollama.com). - Start Ollama with a bigger context. Run
OLLAMA_CONTEXT_LENGTH=65536 ollama serve, or set context length in the Ollama app’s Settings. - Check it is up. Run
curl http://localhost:11434/v1/models. - Connect in Nod. Go to Settings, Models, Add connection, then OpenAI-compatible. Enter
http://localhost:11434/v1and click Connect.
Other addresses: vLLM uses http://localhost:8000/v1 and llama.cpp uses http://localhost:8080/v1. For source or self-hosted installs, set it in .env instead.
LLM_PROVIDER=compatible
OPENAI_COMPATIBLE_BASE_URL=http://localhost:11434/v1
OPENAI_COMPATIBLE_API_KEY=
LLM_MODEL=gpt-oss:20b
Local usage shows as “No price” and does not count toward your monthly limit. Smaller local models make the trust check less reliable, so you can assign a separate checker model. For model ideas, see best Ollama models.
Permissions, Vault and safety
An agent that can pay and send email needs guard rails. Nod’s design puts the key rules in server code, outside the main model.
When Nod asks before acting
- Your words don’t cover it: the trust check asks whenever an action is not clearly part of your request.
- Risky browser steps: sending, posting, ordering, paying, booking, deleting and saving account changes count as acting. Searching and navigating do not.
- Your rules: a rule such as “Always ask me before you send email outside my company” wins, even over your request.
- Checker failure: if the checker can’t answer, Nod asks. In background work the action simply does not run.
Approval cards show the exact account, recipient or amount. A yes runs that exact action once. Reads and drafts run by default unless your rules say otherwise. Say “Show my rules” in chat to review them.
Vault and data
- Vault: saved logins, codes and cards are encrypted on your machine and typed in through private forms the model cannot read.
- Prompt injection: web pages and emails are marked as data, so they cannot give Nod orders.
- Local data: memory, logs and saved logins stay on the machine running Nod. Your provider gets only what each task needs.
- Audit log: the Usage page records every sign-in, card use, action and cost, with CSV export.
Note: app connections go through a connection service that n8n hosts by default. n8n is also open about weak spots: the trust check can be wrong, and it cannot cap a card amount at the issuer.
What Nod costs
- Monthly limit: Nod sets a $200 limit when you finish setup. Change it in Usage (Cmd+5 on Mac, Ctrl+5 on Windows or Linux).
- Task check-in: Nod checks in after $5 of spend since your last message. Say “Check in when a task passes $20” to change it.
- Receipts: each task shows its cost. The log lists every model call with tokens and price.
- Schedules: each recurring schedule shows a projected monthly cost.
- Not counted: ChatGPT plan usage, free models and local servers have no price, so they skip the limit.
Some early testers found tasks cost more than expected. n8n says later releases cut costs a lot. Trust checks also count toward the limit, and an OpenAI key uses a paid checker model separate from any ChatGPT plan.
Research preview limits
- Sleep stops it: schedules run only while the machine running Nod is on.
- One owner: each install supports one person. Team sharing is on the roadmap.
- No phone calls: Nod can’t make calls yet.
- No hosted version: it is planned, with no date.
- No custom skills yet: also listed as “next on the list”.
- Fast updates: there were five releases from v0.7.1 to v0.7.5 on 12-13 November alone, so expect rough edges.
- Long chats: very long chats can hit the model’s context limit, a known issue being worked on.
Nod vs Claude Cowork, Dots, Perplexity Computer and OpenClaw
Here is how Nod stacks up against the other big personal agents in late 2026. Plans and prices change often, so check each vendor before you pay.
| Agent | Where it runs | Models | Price | Open source |
|---|---|---|---|---|
| Nod (n8n Labs) | Your Mac, PC or server, in a local sandbox and browser | Any: Anthropic, OpenAI, Gemini, OpenRouter, 20+ more, local | Free app, you pay API usage | Yes, Apache 2.0 |
| Claude Cowork | Claude desktop app on Mac and Windows, with access to local files | Claude models only | Included on paid Claude plans | No |
| ChatGPT Dots | OpenAI’s cloud computer and browser | OpenAI models (GPT-6 Astra) | ChatGPT Pro and Business Premium | No |
| Perplexity Computer | Perplexity’s cloud sandbox | Perplexity routes across many models | Pro buys credits, Max $200/month with credits | No |
| OpenClaw | Self-hosted Node.js gateway on your hardware | Bring your own keys, local via Ollama | Free, you pay API usage | Yes, MIT |
Which one should you pick?
- Pick Nod if you want model freedom, local data and hard spending limits without running servers.
- Pick Claude Cowork if you already pay for Claude and mostly work with local files. See what is Claude Cowork.
- Pick Dots or Perplexity Computer if you want an always-on cloud agent and don’t mind vendor hosting and plan prices.
- Pick OpenClaw if you want 20+ chat channels and are happy to handle firewalls, patches and security yourself.
My Take
Nod is one of the most thoughtful open-source agent launches this year. The trust check, the Vault and the default $200 cap answer the exact fears people have about letting an AI pay or send email.
It is still a preview, though. The sleep limit, the WSL 2 setup on Windows and the frequent updates mean it suits tinkerers more than busy non-technical users today.
My advice: start with low-risk jobs like a morning brief or price watch, use a strong model, and add a rule that it always asks before payments.
Troubleshooting: Common Problems and Fixes
| Problem | Fix |
|---|---|
| macOS blocks the first launch | Install the newest release and click Open at the Gatekeeper prompt. On a managed Mac, ask IT to allow Nod. |
| Setup fails | Click Try again. If it fails again, use Send feedback from the menu bar or tray. |
| Nod asks for a runtime file | Click Choose runtime file and pick the nod-Runtime-... file from the same release. |
| Windows can’t download the runtime | Put nod-Runtime-win32-x64.tar.gz next to nod-Setup-x64.exe and run the installer again. |
| Windows repeats the same update | Builds 0.6.2 and older can’t self-update on Windows. Quit Nod from the tray and run the newest installer once. |
| Scheduled task didn’t run | Make sure the computer was awake and Nod was running. |
| Local model ignores requests or forgets tools | The context is too small. Raise it to 65,536 or more and reload the model. |
| Long chat fails even on short messages | Update Nod (0.7.5 improved context recovery) or switch to a model with a bigger context window in Settings, Models. |
| Provider rejects a listed model | A saved key doesn’t prove your account has credit or access. Check your provider billing, then click Reload models. |
Warning: don’t use Remove all data as a quick fix. It deletes your conversations, workspace files and saved keys.
Frequently Asked Questions
Is Nod by n8n really free?
Yes. Nod is free and open source under Apache 2.0. You pay only your AI model provider, or nothing per request if you use a local model.
Do I need an n8n account to use Nod?
No. Nod is a standalone app and needs no n8n account or server. You can optionally connect an n8n instance so Nod can work with your workflows.
Does Nod work with Claude and Gemini?
Yes. Anthropic, Google Gemini, OpenAI, OpenRouter, Bedrock, Vertex AI, Azure and over 20 other providers are supported. n8n recommends Claude Opus 5.5 at high effort for Anthropic connections.
Can I use my ChatGPT Plus subscription instead of an API key?
Yes. You can sign in with a ChatGPT Plus or Pro plan in Settings, Models. Plan usage does not count toward Nod’s monthly spending limit.
Does Nod support Ollama and local models?
Yes. Connect any OpenAI-compatible server, such as Ollama at http://localhost:11434/v1. The model must support tool calls and at least 64,000 tokens of context.
Is there a Linux version of Nod?
There is a Linux x64 Flatpak preview, and you can install from source or self-host on a Linux server. ARM Linux builds are not published.
Is it safe to let Nod make payments?
Nod asks before payments you did not request, and card details stay in an encrypted Vault the model cannot read. The trust check can still be wrong, so supervise it during the preview.
Sources
- n8n Blog: Meet Nod
- GitHub: n8n-io/nod
- GitHub: Nod releases
- Nod docs: Install the desktop app
- Nod docs: Connect a model service
- Nod docs: Run a model locally
- Nod docs: When Nod asks before acting
- Nod docs: Spending and usage
- n8n Community: Nod research preview announcement
- Anadolu Agency: OpenAI unveils Dots
- The Decoder: Claude Cowork on all paid plans
- DataCamp: Perplexity Computer
- Fast.io: OpenClaw open source licensing



