中
◀ Back to feed
n8n 官方博客

Meet Nod: a standalone personal agent from n8n Labs that runs on your computer

Published 2026-11-10 Collected 2026-11-11 12:00 AI news #ai#enterprise
Full text

Today we're releasing Nod as a research preview from n8n Labs, a new team within n8n that experiments with ideas outside our core product. Nod is a standalone app, separate from n8n, that runs on your computer. It's a personal agent with its own computer and browser: you tell it what you need, and it works in your apps and on websites until the job is finished. It's open source, and it uses the AI provider you choose.

You can download Nod for Mac and Windows. The code is on GitHub under an open-source licence.

Why we built it

n8n has always been about helping people automate work. That was true before AI and it's true now. AI lets software take on more kinds of work, but the goal is the same.

A lot of what people build in n8n is multiplayer: processes that a team or a business shares and runs on a shared n8n instance. Nod is for a single player. It's built for one person, connects to that person's accounts, and gets their jobs done, whether the job is for their business or their household.

Nod also lives outside n8n. It's a separate app from n8n and from the agents you can build inside n8n, and you don't need an n8n account to use it. You install it on your computer, connect your accounts and ask for something. There's no server or cloud service to set up first.

When we sat down with early testers, many of them placed Nod in the same spot on their own: their chat assistant is where they think and write, n8n runs their shared processes, and Nod takes the ongoing jobs that are theirs alone.

Nod is a n8n Labs experiment, and it doesn't change the direction of n8n. Work on n8n and the n8n agent continues as planned. Building and using Nod has taught us a lot about how AI can take on more of the work people do, and we're releasing it as a research preview so more people can use it and show us where it should go next.

What Nod does

Because Nod has its own computer (a sandbox that runs locally) and browser, it can finish jobs that a chat tool would hand back to you as a to-do list. It works in the apps you connect, such as Gmail, Calendar, Drive and Slack, and on websites that have no integration at all.

Some of the jobs people hand it:

  • Cancel a subscription, turn down the retention offers and ask for your refund
  • Watch a price and tell you when it drops
  • Find a hotel that fits your trip and book it once you say go
  • Sign in to your supplier portals each month, download the invoices, and file them for your accountant
  • Pull last week's numbers from a dashboard that has no way to export the data and put them in your weekly report

Nod also works ahead of you: it reads your connected apps and offers the jobs it could take off your plate. You talk to it in one conversation, in the app or on Telegram or WhatsApp, while it runs several jobs in parallel in the background. Jobs that repeat, like a morning brief or a price check, can run on a schedule.

It gets to know you

Nod remembers months of conversation, so you don't have to keep reminding it what you want. On LongMemEval, a public benchmark for long-term memory, it scored 93.2%, among the best published memory systems.

It learns your preferences as you go, and when you connect Gmail it checks how you write so its drafts sound like you. It also writes down how it got a task done, uses those notes the next time, and fixes them when a site changes. You can see, edit or delete anything Nod keeps about you.

Who it's for

Nod is for anyone who wants to hand the mundane parts of their week to AI: the forms, the follow-ups, the logins and the reminders. Solo operators and freelancers use it for their business, and plenty of people use it for errands at home.

For now you connect your own AI provider, with an API key or your ChatGPT plan, so Nod suits people who are comfortable with that kind of setup. If you use n8n, you can connect your instance so Nod can work with your workflows.

Built to be trustworthy

An agent that can send email and pay with your card has to earn your trust. We designed Nod so you can check how it behaves.

It's yours

Nod is open source under Apache 2.0, and every rule it follows is in the code. It runs on your Mac, your Windows PC or your own server, with the model provider you choose.

It works in its own sandbox

Nod works inside a private virtual machine on your computer, with its own browser. Its data, memory and saved logins stay on the machine running it; your AI provider gets what each task needs.

Only you give it permission

Before Nod sends, pays, posts or deletes anything you didn't ask for, a separate check confirms you asked for it. A web page or an email can't give Nod orders.

The AI never sees your passwords

Your saved logins and cards are encrypted on your machine, and Nod enters them through private forms the model can't read.

Everything is on record

Nod logs every sign-in, card use and action in your apps, along with what each task cost.

The permission check is itself a model and can make mistakes, which is why Nod asks when your words don't cover an action.

What it costs

Nod is open source and free. You pay your model provider for the AI you use, or you can connect an open-source model running on your own computer. Some early testers found tasks cost more than they expected, so cost has been a focus of our recent releases, and the latest one brought it down substantially.

Nod also has guardrails so costs don't run away:

  • Every task shows what it cost, and a finished task ends with a receipt
  • You can set a monthly spending limit
  • Nod checks in with you once a task passes $5, and you can change that amount
  • Each schedule shows its estimated monthly cost

What it doesn't do, yet

  • When Nod is running on your laptop, it stops when the laptop sleeps. Some testers keep it on an always-on Mac mini or run it on their own server, and a hosted version is on our roadmap.
  • Each installation has one owner.
  • It can't make phone calls.

Try it

Download Nod for Mac or Windows, or read the code on GitHub. Then hand it one job from your list. Tell us how it went in the n8n Labs category on the community forum, or with Send feedback in the app.

More info:

https://github.com/n8n-io/nod - Github repo

https://nod.n8n.io/docs/ - Nod Docs

Download and install Nod:

https://nod.n8n.io

Source
Title
Meet Nod: a standalone personal agent from n8n Labs that runs on your computer
Author
Ophir Prusak
Site
n8n 官方博客
Published
2026-11-10
Collected
2026-11-11 12:00 · 2026-11-11-12-ai.json
Link
https://blog.n8n.io/meet-nod/
Open original ↗

Reposted. Will be removed on request. All rights belong to the original author.

More from this batch

12:00batch · 2026-11-1112 items
>

阿里发布 Qwen4-Max 预览版:主打编程和智能体,先在 Token Plan 和自家产品里开放

Qwen 官方只说新一代在知识、世界理解、编程和智能体任务上更强,面向复杂编程、前端搭建和专业软件操作;目前没公布参数、跑分、按量价格和正式发布日期,模型 ID 是 qwen4-max-preview。可以通过千问 Token Plan(限时最低 1 折)接到 Claude Co…

凤凰网科技
>

Nvidia hires law firm Wachtell Lipton to lead Jensen Huang affair probe

On Sunday, Huang admitted to having an affair with a woman who has been connected to Chinese intelligence by reports in The New York Times a…

CNBC
>

Anthropic opens Australian retail IPO offering as AI scrutiny intensifies

Anthropic opened its Australia retail IPO offering on Wednesday, allowing individual investors to apply for shares ahead of its planned IPO.

CNBC
>

CoreWeave 三季度营收 36.6 亿美元、同比涨 168%,上调全年指引,但到 2027 年不再发母公司无担保债

收入和指引都超预期,说明 AI 算力需求没降温;但股价过去一个月跌了四成多、部分债券收益率超 16%,CFO 表示按当前利率不再发母公司层面无担保债,改用客户预付款、硬件伙伴融资和合同挂钩的债务。新云厂商的瓶颈已经从需求转到资金成本。

CNBC
>

亚马逊 Ring 推出 AI 巡逻无人机和地面机器人,2026 年底到 2027 年初上市

Ring Rover Cam 和 Ring Drone Cam 用 AI Pilot 系统结合视觉模型、端侧传感器和云端处理规划巡逻路线,并自动绘制家里或店铺的平面图、建议巡逻点。会动的摄像头也带来新的隐私问题,Ars 回顾了亚马逊上次推室内飞行摄像头时引发的争议。

Ars Technica
>

Tendrils 获 Benchmark 领投 3000 万美元,要从头重做 CPU:数千核并行、配套开源语言 Vine

他们认为现有 CPU 被向后兼容拖累,卡在缓存一致性和乱序执行上,于是放弃共享可变内存和线性指令流,做并行优先的处理器;英国主权 AI 基金、Conviction 等参投,Jack Altman 进董事会。离量产还很远,但代表了 AI 时代对通用计算架构的又一次押注。

Tendrils
>

删权重就能后训练:只把不到 0.05% 的权重置零,效果追平全量微调,还几乎不遗忘

进步在哪:在可验证奖励强化学习、监督微调和在线蒸馏三种后训练里,Deletion Tuning 都达到全量微调的水平,训练出的掩码比完整检查点小几百到几千倍;强化学习时旧任务准确率基本保持,而全量微调、LoRA 会掉最多 5.4 分。靠什么:只在 Transformer 门控投影…

arXiv 2611.10866
>

Meta FAIR 发布 e-means:k-means 的直接替代,约十分之一的轮数达到同样损失,已进 FAISS 1.16.0

进步在哪:在 24 种数据配置上,1.4 到 2.6 轮就达到 Lloyd k-means 跑 25 轮的损失,GPU 上比 Flash-KMeans 快 10 到 45 倍,测试规模到 100 亿点、1670 万簇。靠什么:小批量更新,用指数滑动平均让旧的统计量按比例衰减,而不…

MarkTechPost
>

TIS-100 编程谜题评测:16 个模型在 55 道题上找到 133 个新的帕累托前沿解,3 项周期数纪录被刷新

TIS-100 是一个极度受限的汇编小游戏,玩家社区优化了十年。作者给模型游戏和手册、不给排行榜,让它们不断改进;Opus 5.5 和 GPT-6 Astra 领先、全部解出,几天后人类玩家又在模型解法上继续改进。说明模型在长时间、多目标的代码优化上已经能做出人类没想到的东西。

TIS-100 Bench
>

dNaViT:离散编码也能做通用视觉表示,一个分词器同时支撑理解、生成和编辑

进步在哪:原生分辨率的离散视觉 Transformer,在理解、生成、编辑任务上追平或超过专用基线。靠什么:分组乘积量化加双分支训练目标,让离散 token 同时保留语义和重建细节。对做统一多模态模型(同一套 token 既看又画)的路线是个支撑。

arXiv 2611.12426
>

世界模型要见几次才学会新规则?在测试时学习没见过的游戏机制

作者把每次观察到某条机制定义为一次「曝光」,系统测量世界模型需要多少次曝光才能预测新机制的效果。这关系到智能体进入陌生环境(比如新软件、新业务系统)时,能多快靠少量交互学会规则。

arXiv 2611.12324
>

World ICL:让机器人在上下文里学操作技能,而不是把技能写死进权重

进步在哪:以往上下文学习在机器人操作上一直打不过训练出来的策略,这篇尝试缩小差距。靠什么:把「世界」本身作为上下文示例给智能体,让它在部署时从示范中学技能,避免训练分布绑定。

arXiv 2611.12885