内容来源：https://blog.n8n.io/introducing-n8n-assistant/

内容总结：

n8n推出AI助手，自然语言即可构建自动化工作流

知名自动化平台n8n近日发布全新功能——n8n Assistant，用户只需用日常语言描述自动化需求，即可自动完成工作流的规划、构建、调试和优化，大幅降低自动化项目的搭建门槛。

长期以来，许多用户在n8n平台上常常面临“说得清需求，却不知道如何配置节点”的困境。从描述想法到实际搭建工作流之间的“翻译”过程，往往耗费大量时间，也令不少项目中途搁浅。

n8n Assistant正是为解决这一痛点而设计。用户只需描述期望的结果，它便会自动规划并搭建工作流，在需要时主动询问并引导连接账户凭证，随后执行工作流。一旦遇到节点报错，它能够自主排查原因、提出修复方案并迭代优化，直至任务成功完成。

这一过程完全在n8n现有画布上运行，生成的并非一次性代码，而是团队可长期拥有和运维的标准工作流：每个节点清晰可见，执行历史全程记录，手动可编辑。与市面上其他仅能生成代码片段的AI工具不同，n8n Assistant生成的成果不会随任务结束而消失，而是能够持续稳定运行，也方便同事接手理解。

核心功能亮点包括：

• 先问后做：需求不明确时主动追问，避免错误构建；
• 画布搭建：节点自动连线，随时手动调整；
• 动态验证：实际运行测试，而非空谈成功；
• 自主调试：分析失败原因、修复并重新执行；
• 关键操作确认：涉及凭证权限和工作流激活均需用户明确同意。

据悉，n8n Assistant的典型应用场景包括：表单提交自动分派、CRM数据定期同步、生成每日失败执行汇总，以及修复长期静默出错的既有工作流。

适用范围与费用： 该功能已面向n8n Cloud新实例默认开启，企业云实例暂未包含；自托管Docker用户需自备API密钥并配置环境变量（版本需为2.36及以上），npm方式暂不支持，企业版正在开发中。使用量会计入用户的AI信用额度，现有额度分配不受影响，超额补充功能将在未来几周内上线。

由于该功能目前仍处于预览阶段，n8n团队表示正积极收集用户反馈，特别是在具体使用场景中遇到的“你说什么、它建了什么、哪里出了错”等细节信息，以便持续优化产品体验。

中文翻译：

如果你用n8n构建过任何东西，你就会知道那道坎：你能用一句话描述你想自动化什么，但这和知道该用哪些节点、按什么顺序、如何配置、连接哪个账户完全是两回事。当第一次执行遇到你从没见过的报错时，你又得翻文档、刷论坛了。

大多数自动化项目就是在这一步卡住的。不是因为平台做不到，而是因为琢磨出怎么做要花掉一个你本来没打算搭进去的下午。

有了n8n Assistant，你用日常语言描述你想要的自动化，它会规划工作流、在你的画布上搭建它、向你索要所需的凭据、运行它、找出哪里出了问题，并和你一起迭代修复。

你最终得到的是一个标准的n8n工作流：在画布上清晰可见、可以手动编辑、每次运行都有日志记录、由你的团队真正拥有，而不是由写提示词的人独占。

你会在几分钟内跑通一个能用的东西，然后逐步完善它：替换节点、调整过滤器、添加错误处理、决定哪里需要确认步骤。如果你想自己动手组装，也完全没人拦你。画布是一样的，工作流也还是那些工作流。

现在不缺能从一句话生成自动化的AI工具，其中几个做得还相当不错。但问题在后面才暴露——当团队不得不依赖这个结果时：

• 无从检查。让编码智能体给你写个集成，你得到的是代码，而代码总得有人去托管、去加固安全、去监控、去维护。如果你是个开发者自动化自己的活儿，这没问题。但当结果只有写提示词的人自己能看懂时，这就很难向别人交代了。
• 无法留存。让一个开放式助手去执行任务，它可能确实能完成。但运行一结束，成果也就跟着消失了。没有留下任何可以交接的东西，下星期也没法再跑一次。
• 没法解释。当自动化做出让人意想不到的事情时，总得有人向没参与构建的人解释清楚。“模型自己决定的”这说法糊弄不过去。

所以问题不在于你能不能生成一个自动化，而在于下个月它还能不能跑、你能不能看到它做了什么、同事能不能接手。

n8n Assistant做什么

你从描述想要的结果开始，而不是描述实现步骤：

• 它会在假设之前先提问。如果你的请求含糊不清，它会追问澄清，而不是直接构建错误的东西。“把新的表单提交发给我的团队”会触发一个追问：是哪个表单、发到哪个渠道。
• 它在你的画布上构建。节点出现在你的工作流中，连接妥当，打开即用。没有需要托管到别处的脚本或配置文件。
• 它在需要时才向你索要凭据。你还没看到任何东西跑通之前，不需要先完成一长串设置清单。当它到达一个需要连接账户的节点时，它会当场询问，并引导你完成操作。
• 它会运行工作流。n8n Assistant执行它构建的内容，观察运行过程，读取你本来会自己打开看的那些执行数据。
• 它会调试。如果某个节点出错，它会找出原因、提出修复方案——然后应用修复，再次运行。不需要在工具之间复制粘贴，也不用下载JSON文件人工检查。
• 在所有关键操作之前，它都会先征求你的同意。凭据访问和工作流激活都需要你明确确认。

你的工作流始终是你的

这一切都不是发生在一个独立的产品界面里。n8n Assistant构建的工作流就是你会手动构建的那个工作流，在同一个项目中，用同样的节点，带着同样的执行历史。

这里有个典型例子。有人想对收到的表单提交进行分类处理：补充公司信息、检查是否已是客户、然后要么转给销售要么发送模板回复。这些对他们来说只是一句话的事，但画成节点图就完全没概念了。n8n Assistant规划好方案、搭建工作流、在到达需要CRM凭据的节点时提出请求、用一条测试提交运行，然后发现信息补充这一步对个体经营者返回了空数组。它修复了这个问题，再次运行，最后交付一个有八个节点和一次干净执行记录的工作流。

一周后路由规则变了。一个没参与过的人打开工作流，读了一遍，编辑了一个节点。他不需要知道当初谁输入了什么提示词才建出这个工作流。

运行它、审查它、改进它

生成工作流只是第一步。之后你得到的是：

• 它真的会执行。n8n Assistant运行工作流并读取结果，而不是直接宣布成功。
• 失败可见。执行日志按节点显示输入、输出和错误，和手动调试时看到的界面一样。
• 修复会被应用并重新运行。这个循环一直持续到工作流跑通，或者卡在只有你能解决的障碍上——比如一个还不存在的账户。
• 确认关卡始终在你手里。凭据访问和激活都在等你批准。
• 在依赖它之前先审查。阅读工作流、检查凭据、查看日志，然后再激活。尤其是对于任何业务关键型的内容。

一次构建要花多少

n8n Assistant消耗你套餐中的AI积分配额。现有用户已有的配额不会减少，使用量与此前的AI工作流生成器分开计算，所以你在新功能上的消耗不会影响你原有的额度。

需要多轮调试的构建比一次跑通的更耗积分。积分按套餐分配，我们将在未来几周增加充值方式。

可用性

• n8n Cloud：新实例默认开启。企业版Cloud实例暂不纳入此默认范围。
• 自托管（Docker）：你自带密钥，设置额外的环境变量，运行2.36及以上版本的n8n。详见 https://docs.n8n.io/deploy/host-n8n/install-options/one-line-setup
• 自托管（npm）：暂不支持。
• 企业版：已列入路线图，即将推出。

请注意，n8n Assistant目前仍处于预览标志之后。这不代表使用人数受限：它正在积极开发中，更多功能即将上线。

有些东西还没有，但我们在为未来努力：

• 第一个工作流不保证达到生产就绪标准。它帮你跑通一个能用的东西，但审查仍然是你的工作。
• 凭据和第三方配置的麻烦没有消失。n8n Assistant会在合适的时机询问并引导你，但它没法在你没有账户的服务上帮你注册一个。
• 它不是主动式的。它不会监控你的实例、随时间学习你的偏好，也不会推荐你没要求过的自动化。
• 它只作用于单个实例。本次发布不包含多实例能力。
• 它不会操控你的电脑或浏览器。浏览器辅助的凭据设置是我们正在探索的方向，不在此次发布范围内。

另外，n8n Assistant取代了AI工作流生成器。旧版生成器只生成一次工作流就完事了，而n8n Assistant会构建它、运行它，并寻找改进空间。

上手开始

在你的实例中打开n8n Assistant，描述一个你一直拖着没做的自动化。挑一个真实的需求，涉及两三个集成，并且至少有一个环节可能会出问题。

这里有些思路供你参考：

• 把收到的表单提交路由给对应的负责人
• 按计划将电子表格同步到CRM
• 每天把昨天失败执行的摘要发到一个频道
• 把重复性的手工报告变成工作流
• 修复一个一直在悄悄报错的现有工作流

查看n8n Assistant文档：https://docs.n8n.io/build/ways-of-building-workflows/n8n-assistant

由于n8n Assistant还在预览标志后面，最有价值的反馈是具体的那种：你问了什么、它构建了什么、在哪里翻了车。我们正密切关注社区论坛。

我们迫不及待想看到你用它能构建出什么。

英文来源：

If you've built anything in n8n, you know the gap: you can describe what you want to automate in a sentence, but that's not the same as knowing which nodes to use, in what order, configured how, and connected to which accounts. And when the first execution fails with an error you haven't seen before, you're back in the docs and the forum.
That translation step is where most automation projects stall. Not because the platform can't do it, but because working out how takes an afternoon you didn't plan to spend.
With n8n Assistant, you describe the automation you want in plain language and it plans the workflow, builds it on your canvas, asks for the credentials it needs, runs it, works out what broke, and iterates on the fix with you.
What you get at the end is an standard n8n workflow: visible on the canvas, editable by hand, logged on every run, and owned by your team rather than by whoever wrote the prompt.
You'll get something working in a few minutes, then you tighten it: swap a node, adjust a filter, add error handling, decide what needs a confirmation step. If you'd rather assemble it yourself, nothing stops you. It's the same canvas, and the workflows are the same workflows.
There's no shortage of AI tools that'll produce an automation from a sentence, and several of them are fairly good at it. But problems show up later, when a team has to depend on the result:

• Nothing to inspect. Ask a coding agent for an integration and you get code, which somebody now has to host, secure, monitor and maintain. That's fine if you're a developer automating your own work. It's harder to justify when the result only makes sense to the person who prompted for it.
• Nothing that persists. Ask an open-ended assistant to go and do the task and it might well do it. But the run ends, and the work goes with it. There's nothing left to hand over or run again next week.
• Nothing to explain with. When an automation does something surprising, someone has to explain it to a person who didn't build it. "The model decided to" doesn't cover it.
  So the question isn't whether you can generate an automation. It's whether it still runs next month, whether you can see what it did, and whether a colleague can pick it up.
  What n8n Assistant does
  You start by describing the outcome you want rather than the steps to get there:
• It asks before it assumes. If your request is ambiguous, it asks a follow-up instead of building the wrong thing. "Send new form responses to my team" gets a question about which form and which channel.
• It builds on your canvas. Nodes appear in your workflow, wired together, ready to open. There's no script or config file to host somewhere else.
• It asks for credentials when it needs them. No setup checklist before you've seen anything work. When it reaches a node that needs an account connected, it asks then and guides you through it.
• It runs the workflow. n8n Assistant executes what it built, watches what happens, and reads the same execution data you'd open yourself.
• It debugs. If a node fails, it works out why, proposes a fix - then applies it, and runs it again. No copy and pasting between tools or downloading JSON files to manually review.
• It asks before anything consequential. Credential access and workflow activation need your explicit confirmation.
  Your workflows stay yours
  None of this happens in a separate product surface. The workflow n8n Assistant builds is the workflow you would have built by hand, in the same project, with the same nodes and the same execution history.
  Here's a typical example. Someone wants inbound form submissions triaged: enrich the company, check whether they're already a customer, then either route to sales or send a templated reply. They know all of that as a sentence and none of it as a node graph. n8n Assistant plans it, builds it, asks for the CRM credential when it reaches the node that needs one, runs it against a test submission, and finds that the enrichment step returns an empty array for sole traders. It fixes that, runs it again, and hands over a workflow with eight nodes and a clean execution.
  A week later the routing rules change. Someone who wasn't involved opens the workflow, reads it, and edits a node. They don't need to know what anyone typed to get it built.
  Run it, review it, improve it
  Generating the workflow is only the first step. What you get after that:
• It executes. n8n Assistant runs the workflow and reads the result instead of declaring success.
• Failures are visible. Execution logs show inputs, outputs and errors per node, the same view you'd use to debug by hand.
• Fixes are applied and re-run. The loop continues until the workflow completes, or until the blocker is something only you can fix, like an account that doesn't exist yet.
• Confirmation gates stay with you. Credential access and activation wait for your approval.
• Review before you rely on it. Read the workflow, check the credentials, look at the logs, then activate. Especially for anything business-critical.
  What a build costs
  n8n Assistant draws on your plan's AI credit allocation. Existing users keep the allocation they already have, and usage is counted separately from the previous AI workflow builder, so nothing you spend here changes what you had before.
  A build that needs several rounds of debugging costs more than one that lands first time. Credits are allocated per plan, and we will be adding ways to top these up in the coming weeks.
  Availability
• n8n Cloud: on by default for new instances. Enterprise Cloud instances are excluded from that default for now.
• Self-hosted (Docker): You bring your own keys, set additional environment variables, and run an n8n version from 2.36 onwards. See https://docs.n8n.io/deploy/host-n8n/install-options/one-line-setup for more details.
• Self-hosted (npm): not supported.
• Enterprise: on our roadmap, coming soon.
  Please note that n8n Assistant still ships behind a preview flag. That isn't a limit on who can use it: it's in active development with more features coming soon.
  Some things aren't here yet, but we're working on for the future:
• The first workflow isn't guaranteed to be production-ready. It gets you to something that works. Reviewing it is still your job.
• Credential and third-party setup hasn't gone away. n8n Assistant asks at the right moment and guides you, but it can't create an account on a service you don't have.
• It isn't proactive. It doesn't monitor your instance, learn your preferences over time, or suggest automations you haven't asked for.
• It works on one instance. There are no multi-instance capabilities in this launch.
• It doesn't drive your computer or your browser. Browser-assisted credential setup is something we're exploring, and it isn't part of this release.
  Also, n8n Assistant supersedes the AI Workflow Builder. Where the old builder generated a workflow once and stopped, n8n Assistant builds it, runs it, and finds improvements.
  Get started
  Open n8n Assistant in your instance and describe an automation you've been putting off. Something real, with a couple of integrations and at least one step that could plausibly go wrong.
  Here are some ideas to get started with:
• Route inbound form submissions to the right owner
• Sync a spreadsheet to a CRM on a schedule
• Post a daily summary of yesterday's failed executions to a channel
• Turn a recurring manual report into a workflow
• Repair an existing workflow that's been failing quietly
  Check out the n8n Assistant documentation here: https://docs.n8n.io/build/ways-of-building-workflows/n8n-assistant
  While n8n Assistant is behind the preview flag, the most useful feedback is the specific kind: what you asked for, what it built, and where it fell down. We're reading the community forum closely.
  We can't wait to see what you build with it.