n8n Labs Releases Nod, an Open-Source Personal Agent That Operates in Its Own Computer

Canceling a subscription while declining retention offers, fetching monthly invoices from vendor portals, preparing for a client call: these are the kinds of chores that n8n, the developer of the popular automation tool, is now assigning to an agent. Its experimental team, n8n Labs, has released Nod, an open-source (Apache 2.0) personal agent that runs on the user’s machine, complete with its own virtual computer and browser.

A “Solo” Agent, Separate from n8n

The launch post, authored by Ophir Prusak, emphasizes the distinction. While n8n is used to automate processes shared across a team, Nod is designed for a single individual, connected to their accounts, for both professional and household matters. It is a standalone application that does not require an n8n account (though users can connect their instance to trigger workflows), and n8n makes clear that this project does not change the direction of its core product.

In practice, Nod runs on Mac (macOS 14 or newer), on Windows 11 in preview, or on a server. It acts within connected apps like Gmail, Calendar, Drive, Slack, or Notion, but also on websites that offer no integrations at all by navigating them using its own browser. Users can communicate with it inside the app, via Telegram, or on WhatsApp; it handles multiple tasks in parallel and can schedule recurring routines, such as price tracking or a daily morning briefing. It maintains several months of conversational memory that users can inspect and edit, and it records how it successfully completed a task so it can repeat it later.

Choice of Model, Choice of Bill

Nod is free, but the AI is not: users plug in an OpenAI, Anthropic, or OpenRouter key, their ChatGPT Plus or Pro subscription, or, according to the blog, a locally run open-source model. n8n acknowledges that early testers found tasks more expensive than expected and has added cost safeguards: displaying the cost for each task, setting a monthly cap, and requesting confirmation whenever a task exceeds $5.

Security: The Promise and Its Limits

The most closely scrutinized aspect remains trust. Nod operates within a local virtual machine; its data, memory, and stored credentials stay on the machine running it, and the AI provider receives only what each task requires. Passwords and credit cards are encrypted and entered through forms the model cannot see. Before sending, paying, publishing, or deleting anything that was not explicitly requested, a separate verification step ensures the user actually intended it: according to n8n, an incoming email or a webpage cannot issue commands to the agent. The developer admits, however, that this verification step is itself handled by a model, which can make mistakes.

n8n highlights in-house benchmark scores: a 98% success rate on 226 test tasks, 93.2% on the LongMemEval-S memory test, and zero false memories implanted across 216 attempts using prompt-injection emails (WhisperBench), results achieved with GPT-6.1 Sol. No independent benchmarks are available yet. Other acknowledged limitations: the agent stops when the laptop goes to sleep, each installation supports only a single owner, and it cannot make phone calls. A hosted version is planned.