Newshot n8n Launches Local Agent Nod: It Books Hotels, and Asks Before Paying
Nod is a local research preview personal agent from n8n Labs that can perform tasks such as booking hotels and canceling subscriptions across apps and websites. It limits risk through upfront confirmation, a local sandbox, and an action log, but it stops when the laptop sleeps, is single-user only, and costs may exceed expectations.
- Nod is open source as a research preview
- Runs in a local sandbox and connects to personal accounts
- Checks separately before actions such as sending and paying
- Monthly limits and task receipts control costs
This article is based on the release post on the official n8n blog; features, pricing, and test results all come from the vendor's own statements, and no independent retesting is provided in this article.
n8n Labs has put the personal agent Nod on your own computer: it can book hotels, cancel subscriptions, and download invoices across apps and websites, and when it encounters actions such as paying, sending, publishing, or deleting, it stops first to check against your own words. The software is open source and free, model costs are separate, and any single task over 5 USD requires your approval by default. Whether you can confidently hand it the payment button depends not on whether it can search for hotels, but on whether, each time it stops, you can understand what it is doing and how much it has spent.
It doesn't live in the cloud; it lives in your laptop
Nod is a personal agent from n8n Labs. It is independent of the core n8n product, installs on your own computer, and connects to your personal accounts—not the shared workflows used by a team.
It handles "one person's to-do list" and is not aimed at team-shared workflows. Once Nod is installed, accounts are connected, and you have stated what you want, it goes into apps and web pages and gets the job done. At this stage it is released as a research preview (a test version not yet formally finalized), and the code uses the Apache 2.0 open source license; as a standalone app, it does not require an n8n account to use.
It can get into apps with official connections such as Gmail, Calendar, Drive, and Slack, but the other half is what really sets it apart: on websites without official connections, it still clicks, fills in forms, and turns pages in the browser, carrying a whole task from start to finish rather than just pressing a ready-made button inside some app.
The officially listed tasks are very specific: find hotels that match travel needs and book them after the user confirms. It can connect search with follow-up action, but booking still depends on the user's approval, and it will not pay directly just because it sees a suitable result.
How does it complete hotel bookings on the user's behalf?
After receiving a booking task, Nod completes the reservation in its own browser rather than handing the page back to you. When it encounters actions such as sending, paying, publishing, or deleting, it first checks against your own words.
This way of executing comes from Nod's local sandbox (an isolated runtime environment inside the computer) and a separate browser. It can work in connected apps and also operate websites without dedicated interfaces, so the to-do items that chat tools originally listed for you can be picked up and carried forward by it. The key here is not giving longer answers, but crossing from "finding a hotel" to actually booking.
You can also talk to it via Telegram (a messaging app) or WhatsApp (a messaging app) without staying at the computer; the browser continues searching and filling in pages in the background.
Multiple tasks can run in the background at the same time, so you do not have to wait for a hotel booking to finish before starting the next thing; tasks that need to be repeated can also be started on a schedule, such as checking prices at set times. Running tasks in parallel solves the waiting problem; separate checks solve the permission problem.
The check itself is another model, and it can also make mistakes. So its criterion is very narrow: it only recognizes whether your words cover this action. When booking a hotel, payment is an action that requires special verification; if your words do not explicitly cover it, it stops and asks. This mechanism cannot replace the user's own review of amounts and action logs, but it can add a system-enforced threshold before money is deducted.
It remembers what you said months ago
Ordinary chat assistants give advice, draft emails, and list to-dos, but in the end the user still has to execute each item. Nod enters connected apps and websites and keeps operating until the task is complete.
Nod reads connected apps, proactively identifies work it can take over, and offers to complete it on your behalf. What it faces is a task that needs to be carried out, and the sign is direct: the relevant actions have been completed.
This delegated capability is built on a continuous relationship. Nod remembers conversations from months ago, so users do not have to repeatedly restate preferences and context. On LongMemEval (a long-term memory benchmark used to test whether agents can remember past information), Nod's officially published score is 93.2%.
This is a benchmark result provided by the vendor, and no third-party retest has been seen so far, so it cannot be used alone to judge the actual experience.
After connecting Gmail (Google email), Nod also checks how the user phrased things in the past, making new drafts closer to the user's own tone. An ordinary assistant may temporarily imitate a writing style, while Nod learns from historical expression. This affects whether the emails it writes on the user's behalf can truly represent the user.
After a website is redesigned, tasks also cannot stop halfway. Nod records how a task was completed, reuses those steps next time, and corrects the method when the website changes. The final link in delegation is to keep doing the work, not to hand the matter back to the user.
Our judgment: memory is the prerequisite for its continuous delegation, and beyond that prerequisite one more step of verification is missing—the vendor's benchmark score has no third-party retest, and the actual experience must be tested with real tasks.
Login information stays on the local machine; the model cannot see passwords or cards
Nod has to log in, pay, and send emails for you, so will passwords and card numbers be seen by the model? n8n Labs' answer is no: credential storage, task execution, and model calls are split into three layers, and passwords and cards always stay on your machine.
The first layer is the virtual machine. Nod runs in a private virtual machine on your computer, with its own browser, and logging into websites and executing tasks are all completed inside it, without taking over the browser window you are using. Data, memory, and saved login information all remain on the local machine running Nod and are not centralized in n8n's cloud.
The second layer is encryption. Saved login information and cards are encrypted locally, and submission goes through a private form that the model cannot read. The model only knows "this step requires login and filling in information" and cannot see the specific contents of passwords and cards; the one actually doing the filling is still Nod on the local machine.
The third layer is permissions. Web pages or emails cannot give commands to Nod, and text on the page will not be treated as your new instructions. Before Nod sends, pays, publishes, or deletes, a separate check confirms that these are your words—this check is itself also a model and can also make mistakes, so when it is unsure it comes back to ask you.
Can users see what it did and how much it spent?
Nod puts actions and costs in the same record: users see not only that the task is complete, but also where it logged in, which card it used, and how much the task cost.
According to n8n Labs' official introduction, Nod records every login, card use, and in-app action, while also saving the cost of each task. What is recorded is "what was done" and "how much was spent," so users can review by task instead of only guessing from the final result what the agent did.
Cost control is divided into two layers: routine account viewing and interception before payment. Each task shows its cost, a receipt is generated after the task ends, and users can set a monthly spending cap. Scheduled tasks also show estimated monthly costs, which brings the price of repeated execution into the planning stage in advance rather than raising questions only after costs accumulate.
| Scenario | Prompt or control provided by Nod | What the user can judge |
|---|---|---|
| A single task costs no more than 5 USD | The task page shows the cost, with a receipt after completion | How much this task actually consumed |
| A single task costs more than 5 USD | Confirms with the user before payment; the confirmed amount can be modified | Whether to accept this amount, and whether to adjust before execution |
| Long-term or repeated execution | Shows estimated monthly cost and allows setting a monthly spending cap | Whether ongoing spending remains within budget |
Our judgment: Nod's cost control does not rely on a single master switch, but on records and confirmation flows that lay every expense out in front of the user, on the premise that these records and confirmations are themselves accurate. The official side has not provided third-party retest data, and this is the boundary of the research preview.
Sources: n8n Blog "Meet Nod" and official documentation; features, security mechanisms, pricing, and benchmark results are all based on n8n's own statements, and the research preview has no independent retesting as described in this article.
Glossary · Plain-language notes appearing in this article
Full articles on deep dives, research analysis, and tool tutorials, plus all podcasts, are in the members area.