How to Use n8n Assistant: Build Workflows From Plain Language

Introduction

The blank canvas is the slowest part of automation: you know the workflow you need, but translating it into nodes, parameters, and expressions takes an hour of clicking. n8n Assistant is n8n’s answer, a chat-based agent inside the editor that builds workflows from plain language. This tutorial walks the full loop: writing a prompt it can act on, reviewing its plan, connecting credentials safely, testing before production, and controlling credit spend.

Quick answer

Open the n8n Assistant panel, pick a project, and describe the trigger, apps, data, and success and failure behavior in plain language. Review its plan, connect credentials through the chat card, then test the generated workflow before publishing.

Evidence

n8n launched the AI Assistant on July 9, 2026, in Preview on n8n Cloud from version 2.29.9. The launch announcement frames it plainly: “We’re launching the n8n AI Assistant, a new chat-based agent that lives right inside n8n and builds workflows for you from plain language.” The same announcement says it can “plan the workflow, build it in your selected project, test it, and help you fix errors,” and that the output is a standard workflow you can open, edit, and publish like any other.

An AI builder is only as good as the node ecosystem it can reach. n8n’s GitHub repository had 206,785 stars when ToolPilot checked via the GitHub API on October 9, 2026, which is why the Assistant can draw on thousands of node types and integrations.

The official n8n Assistant documentation confirms availability on Cloud Starter and Pro and on self-hosted Community, Registered Community, and Business editions, while Enterprise is not included yet. The launch announcement states the review rule every user needs: “The result is a normal n8n workflow. You can open it, inspect it, edit it, test it, and publish it like any other workflow.”

Pricing is token-based, not message-based. On Cloud, the Assistant consumes Assistant credits based on the tokens the underlying model processes, with a monthly allowance that resets on the first day of every month and an option for the instance owner to buy extra credits during the Preview. Self-hosted instances skip Assistant credits entirely and run on your own LLM provider account. The separate self-hosted setup thread lists the real requirements: n8n 2.29.7 or later, an API key from Anthropic, OpenAI, or OpenRouter, an optional but recommended search provider (Brave Search or SearXNG), and a required sandbox provider, with Daytona recommended since it is what n8n uses on Cloud.

Prerequisites

• An n8n instance with the Assistant enabled. On Cloud: Starter or Pro plan, version 2.29.9 or higher. On self-hosted: Community, Registered Community, or Business edition, version 2.29.7 or later, plus your own LLM API key and a sandbox provider. Our Docker Compose setup guide covers the self-hosted base.
• Access to the project where the Assistant should work, with permission to use the workflows, credentials, and resources the task needs. The Assistant inherits your permissions and can only touch what you can touch.
• A clear idea of the workflow you want. The Assistant plans from your description, so a vague request produces a vague plan.
• For the examples below: a Gmail account and a Google Drive account if you want to follow the invoice workflow end to end. If you are new to n8n’s AI features, read our n8n Agents tutorial first for the agent vocabulary this article uses.

Step-by-step guide

Step 1: Open the Assistant and choose your project

On n8n Cloud, select AI Assistant from the left sidebar of the editor. This is the documented entry point from the launch announcement, and it opens a chat panel docked next to the canvas so you can watch what the Assistant builds live.

Choose the project before you type anything. The Assistant’s instance context is scoped to that project: workflows and their publication status, recent changes from the activity log, and execution history. That context is what lets you say “debug the latest failed execution” without naming a workflow. Start a new conversation for unrelated work so the Assistant reads the new project from scratch.

Self-hosted users get the same panel once the instance is configured with an LLM provider and a sandbox, via environment variables (a Docker Compose example is in the docs). Expect this part to feel manual: it is a pre-release preview.

Step 2: Describe the task like a specification, not a wish

Write the prompt like a ticket for a colleague. The docs say to include the trigger, the apps or services, the data to read, send, or update, the success behavior, the failure behavior, and whether it should ask before publishing.

A prompt worth copying, adapted from the official examples:

Create a workflow that checks Gmail every morning for invoices,
saves PDF attachments to Google Drive, and adds a row to a Data Table.
If the Gmail search returns nothing, do nothing and log that.
If any step fails, send the error to our #ops Slack channel.
Ask me before publishing the workflow.

Notice what is in there: trigger, apps, data, success path, failure path, and an explicit publish gate. Specific prompts produce better results with fewer iterations, and fewer iterations means fewer credits burned.

Step 3: Review the plan before it builds anything

The Assistant proposes a plan before it creates or changes nodes. Read it. This is the cheapest review point in the whole process: catching a wrong trigger, a missing branch, or a misunderstood app here costs one message, while catching it after the build costs a rebuild.

Check three things in the plan. First, the trigger and schedule, because a workflow that fires on the wrong event is wrong in the most expensive way. Second, the node sequence against your mental model, does it fetch before it transforms, does it filter before it writes. Third, the credentials it intends to use, the Assistant reads credential descriptions (from 2.41.0, and from 2.42.0 only if your instance enables description reading) to choose between credentials of the same type, and it asks you when the choice is unclear. If the plan names the wrong credential, correct it now.

Review the plan before the build, and never ask for a regeneration without new guidance: rebuilding the same wrong answer is the fastest way to burn credits.

Step 4: Let it build, then connect credentials through the chat card

Once you approve the plan, the Assistant builds the workflow in your project, showing its work live. When a node needs credentials, it presents a credential card in the chat: select an existing credential or create a new one.

Enter the secret in the standard n8n credential screen, never in the chat: credential secrets are never sent to the AI, so never paste API keys, passwords, or tokens into the conversation. Keep credential descriptions informative (“production Stripe”) but secret-free.

After the build, open the generated workflow on the canvas and click through the nodes: read the parameters and expressions, and confirm the data mapping matches your source systems. Our AI Agent node tutorial helps if the Assistant built an agent node.

Step 5: Test with real input before anything goes near production

The docs give a five-point review checklist for AI-generated workflows, and it is worth treating as mandatory: review the workflow logic, check node configuration and credentials, test with expected input data, check execution results and error handling, and confirm the workflow does not perform unintended actions.

Run the workflow against a real but harmless input: forward one genuine invoice to a test label and inspect each node’s input and output in the execution. Deliberately trigger a failure (a bad file, a missing field) and confirm the Slack alert fires with the detail you need. The Assistant can run these checks with you and suggest fixes. For systematic patterns, see our n8n error-handling guide.

Only after a clean test run should you let the Assistant publish; it asks for confirmation before high-impact actions like publishing or deleting anyway.

Step 6: Debug failed executions in chat

Because the Assistant reads instance context, you can prompt:

Debug the latest failed execution of this workflow.
Explain what failed, then suggest a fix before changing anything.

The Assistant finds the execution, explains the failure, and only proposes the edit after you approve it, which keeps it in advisory mode during the Preview.

For recurring failures, ask it to harden the workflow instead of patching the symptom: an error branch, a retry with backoff, or a dead-letter path writing failed items to a Data Table the Assistant can create for you.

Step 7: Control credit spend

On Cloud, every message costs tokens, billed as Assistant credits. Four habits keep spend sane: be specific in the first prompt, review the plan before the build, start a new conversation per task, and never regenerate without new guidance.

Check your balance in the credits menu inside the Assistant panel. If the allowance runs out, the instance owner can buy extra credits to reach the next monthly reset. Assistant credits and Gateway credits are separate pools: the first pays for Assistant conversations, the second for the models your workflows call at runtime, and topping up one does not add to the other. On self-hosted there are no Assistant credits, but your own provider bills the tokens, so the same discipline applies.

Step 8: Enable it on a self-hosted instance

Self-hosted setup is officially a manual pre-release process. The requirements from the setup thread: n8n 2.29.7 or later, an API key from Anthropic, OpenAI, or OpenRouter (you bring your own key, your provider bills the usage), an optional search provider (Brave Search or SearXNG, recommended if you want web research), and a sandbox provider, which is required because the Assistant runs its work in an isolated sandbox. Daytona is the recommended sandbox since it is what n8n uses on Cloud; self-hosting n8n’s sandbox service is possible but more involved.

Configuration happens through environment variables, with a Docker Compose example in the docs. Two real community findings matter before you start. First, early preview builds ignored the custom endpoint variable for OpenAI-compatible providers; n8n fixed that in 2.31.0, so use a current release. Second, an openai/ model prefix triggered reasoningEffort: high injection that breaks some third-party models behind custom endpoints; community members report OpenRouter as a working path. If you are still choosing hosting, our VPS install guide covers the base server.

Troubleshooting

The Assistant panel is missing from the sidebar

Check three gates. Plan: Cloud Starter and Pro only, not Enterprise. Version: Cloud needs 2.29.9+, self-hosted 2.29.7+ with the feature configured. Rollout: capabilities land gradually during the Preview, so an eligible instance may simply not have received it yet. On self-hosted, a missing panel usually means the LLM provider or sandbox variables are not set.

Custom OpenAI-compatible endpoint is ignored

Community members reported that a custom model URL was ignored and requests kept going to the default OpenAI endpoint. n8n confirmed the preview bug and fixed it in 2.31.0, so upgrade to 2.31.0 or later. If you pinned an older version for stability, that pin is the cause.

Third-party models fail when configured with an openai/ prefix

With an openai/ model prefix, n8n injects reasoningEffort: high into requests, which some providers behind custom endpoints reject. Users hit this with DeepSeek and Mistral while direct calls without the parameter worked. The community workaround is routing through OpenRouter; otherwise avoid the openai/ prefix so the parameter is not injected.

A generic error appears when credits run out

A July 2026 update to the launch thread notes that exhausted credits currently surface as a generic error, not an out-of-credits message. If the Assistant suddenly fails on a previously working instance, check the balance in the credits menu before debugging anything else. The remedy is a plan upgrade or the owner buying extra credits to reach the next reset.

The Assistant chose the wrong credential, or you pasted a secret into chat

If it picks the wrong credential, add clear descriptions (production versus test, owner team); description reading needs 2.41.0, and from 2.42.0 the instance must enable it. If you pasted a secret into the chat, rotate it immediately and re-enter it only through the standard credential screen.

FAQ

Is n8n Assistant free to use? On n8n Cloud it runs on Assistant credits, not cash per message. Your plan includes a monthly allowance that resets on the first day of every month, and the instance owner can buy extra credits if you run out during the Preview. On a self-hosted instance there are no Assistant credits at all, but you pay your own LLM provider for the tokens the Assistant consumes.

How is n8n Assistant different from the old AI Workflow Builder? The Assistant replaces the AI Workflow Builder. n8n says the Assistant can do everything the builder could do and more, with better results: instead of generating a workflow from a single prompt in a separate window, it is a chat agent that lives inside the editor, plans, builds, edits, tests, and debugs, and shows you what it is doing live.

Can I enter my own LLM API key on n8n Cloud instead of using credits? Not yet. The n8n team says bring-your-own-key on Cloud is something they are thinking about, but it is not available in the Preview. On self-hosted instances you always use your own key from Anthropic, OpenAI, or OpenRouter.

Which model does n8n Assistant use? On n8n Cloud it currently uses Opus 4.8 for most of the work, according to the launch announcement. On self-hosted instances the model is whatever you configure through your own LLM provider account.

Is n8n Assistant available on Enterprise plans? Not yet. It is available on n8n Cloud Starter and Pro, and on self-hosted Community, Registered Community, and Business editions. Enterprise customers on Cloud or self-hosted need to contact their Customer Success Manager about preview access.

Can I edit what the Assistant builds? Yes. Everything the Assistant produces is a normal n8n workflow. You can open it, inspect every node, edit parameters, test it, and publish it exactly like a workflow you built by hand. That is the point: the Assistant is a starting point, not a black box.

Conclusion

n8n Assistant turns the blank canvas into a conversation: describe the automation like a specification, review the plan, connect credentials through the chat card, and test with real input before publishing. Treat it as a fast junior builder and keep the five-point review checklist non-negotiable while the feature is in Preview. Next, our n8n Agents tutorial covers the dedicated agent environment the Assistant can scaffold for you.

Frequently asked questions