Hey community! 
Two months ago we introduced the AI Assistant here on the forum. Since then quite a lot has changed, and we published a blog post this week about what it is and why we built it. So this feels like a good moment for an update: what you can do with it now, and what got better since July.
One small thing first. We’ve renamed it from “AI Assistant” to n8n Assistant. Same product, and you may still see the old name in a few places in the product while we finish the sweep.
1. It does more than build workflows now
In July the list was: create, edit, test, and troubleshoot workflows. That’s still the core, and it’s still the main reason to open it. But the list has grown. You can now ask n8n Assistant to:
- Build agents
Describe the agent you want, and it suggests instructions, tools, and skills to add. Agents are still rolling out, so not everyone has them yet. The agents docs explain how they work and how to enable them on self-hosted. - Connect MCP servers
Connect a server from the MCP server registry directly to the assistant. It can use the server to get things done, and to gather context while it builds a workflow for you. - Create and update Data Tables
When a workflow needs a Data Table, it creates one. You can also edit the table right in the preview panel once the assistant is done working. - Research on the web
When web access is on, it looks up current docs and APIs while building. It asks your permission before accessing a new domain. - Work with folders and your workspace
It can navigate and manage folders, and it looks at the other workflows in a folder to understand the context of what you’re asking for.
2. Fewer credentials to set up
Gateway credits. n8n Cloud recently launched Gateway credits: a prepaid balance you can use to run AI models (OpenAI, Anthropic, Gemini, and others) and some tool services (Brave Search, Firecrawl, and more) without creating a provider account or pasting in an API key.
n8n Assistant knows about them. When it builds a workflow with, say, an OpenAI node and you don’t have an OpenAI credential stored, it sets the node up on Gateway credits so the workflow can run right away. If you do have your own credential, it uses that instead. Same when it builds an agent: Gateway credits show up as a model option next to your own keys. Every Cloud instance gets some free credits to start.
Services without an n8n credential type. This one I’m personally happy about. When a workflow uses an HTTP Request node against a service with no dedicated n8n credential, the assistant now researches how that service does authentication and pre-fills a simple credential form for you (icon, docs link, the fields it needs).
It also checks your key against the service’s API before it saves anything. If the key gets rejected you see a clear error right away instead of a failing workflow later.
The rest of the credential handling works as before: it asks for a credential only when it reaches a node that needs one, and your secrets never go through the chat. You enter them in the normal n8n credential screens.
3. Self-hosting is one line now
If you followed the original self-hosted instructions, you’ll remember they involved a fair number of moving parts. That’s mostly gone. From n8n 2.35, a new Docker install is:
curl -fsSL https://get.n8n.io | sh
That sets up n8n with the code sandbox and web search preconfigured, and n8n Assistant enabled by default. You add your model provider key in the UI (Anthropic, OpenAI, OpenRouter, or any OpenAI-compatible endpoint) and start building. Existing Docker installs update to 2.35 or later and configure the sandbox, search, and model from the new settings page instead of environment variables.
Paul wrote this up in detail in this thread, including what changed for existing installs, so I’ll leave the specifics there. Self-hosted setup questions go in that thread too.
4. The chat and the canvas got smoother
A handful of smaller changes that add up:
- Reasoning shows up where it happened. Instead of one block at the top, each reasoning step appears in order between the tool calls, live and after you reload.
- The logs panel opens by itself as soon as a workflow the assistant built starts executing, so you can watch the data move through the nodes.
- The chat panel doesn’t squash the canvas anymore. It sizes to its own content instead of taking half the screen.
- Connections live under a “+” button next to the chat input, with connected and disconnected status at a glance.
- It stays in your language. There was a bug where it would switch languages mid-conversation. Fixed.
5. It’s more resilient
Conversations now survive restarts, and if a run gets interrupted, the assistant picks it back up rather than losing everything but your first message. We also fixed a few bugs in the build-and-verify loop that could make it pin empty test data or get stuck verifying certain workflows. Less visible than the items above, but if you were hitting those in July, you should notice.
Where things stand
- Still in preview. It can make mistakes and behavior may change. Please review generated workflows before using them in production, especially anything business-critical.
- n8n Cloud: on by default for new instances on Starter and Pro. Existing instances get it when they upgrade.
- Community Self-hosted: Docker only. npm installs aren’t supported.
- Enterprise: on the roadmap, not available yet.
- Credits: n8n Assistant uses your plan’s AI credits. We’re adding the ability to top up AI credits very soon. AI credits and Gateway credits are separate balances: talking to the assistant spends AI credits, and the workflows it builds spend Gateway credits only where a node uses them.
Prompts to try
If you tried it in July and haven’t been back, here are a few ideas that use the newer capabilities:
- “Build an agent that answers questions about our product docs. Suggest the tools and skills it should have before you create anything.”
- “Connect to [MCP server] and use it to check whether these three customers already exist before you build the sync workflow.”
- “Post a daily summary of yesterday’s failed executions to a Slack channel.”
- “Debug the latest failed execution of this workflow. Explain what failed, then suggest a fix before changing anything.”
- “Turn this recurring manual report into a workflow. Ask me before publishing.”
Feedback
The most useful feedback is specific: what you asked for, what it built, and where it fell down. Post it here, and we’ll keep reading.
For self-hosted setup questions, please use this thread so those stay together.
