n8n AI Assistant: building and debugging workflows in plain language

Describe an automation in one sentence and watch n8n build the matching workflow: that's the promise of the AI Assistant, the conversational agent built into the n8n editor since 2026. Unlike a simple autocomplete-style suggestion tool, it can create a complete workflow, edit it on request, run it to spot errors, and propose fixes — all without leaving the n8n canvas or reaching for an external tool. This guide covers what the assistant actually does, how to set it up on Cloud and self-hosted, and the limits worth knowing before handing it a workflow headed for production.

What the AI Assistant actually does

The assistant shows up as a chat panel inside the editor, capable of four kinds of actions:

• Create a workflow from a plain-language description — "when an email lands in this Gmail inbox, summarize it with AI and post the summary to Slack" turns into a trigger, an AI node, and a connected Slack action.
• Edit an existing workflow: add a node, change a condition, adjust a parameter, without rebuilding everything by hand.
• Test and debug: run an execution, inspect the errors a node returns, and propose a targeted fix rather than a generic answer.
• Build an AI agent: suggest system instructions, tools, and skills to wire onto an AI Agent node, in line with what we cover in our full AI Agent node guide.

The important point, worth distinguishing from a simple config copy-paste: the result stays a normal n8n workflow. Nothing is locked away in a black box — every generated node opens, edits, and deletes exactly as if it had been placed on the canvas by hand.

Setting it up on n8n Cloud

On Cloud, the AI Assistant ships in Preview since version 2.29.9, directly accessible from the node search panel or a dedicated icon in the editor — no infrastructure-side configuration is needed, since n8n hosts and manages the model itself. One thing to watch for, flagged by the community: on a trial plan, the assistant can recommend creating an n8n API key even though that option isn't available during the trial period, since the tool has no awareness of your exact plan limits. If a suggestion from the assistant seems to get stuck for no apparent reason, check your plan's restrictions on our n8n Cloud pricing guide before hunting for a configuration error on your end.

Setting it up self-hosted (BYOK)

On a self-hosted instance — the default choice for most readers of this blog, see our self-hosted vs Cloud comparison — the assistant runs on a Bring Your Own Key basis: you supply your own API key and pay the provider directly for usage, with no n8n markup on top. Three things are needed:

1. Enable the module — the N8N_ENABLED_MODULES environment variable must include instance-ai (the module ships enabled by default in recent versions; remove it explicitly from the list if you'd rather disable it).
2. Configure a model provider — Anthropic, OpenAI, and OpenRouter are supported, via the N8N_INSTANCE_AI_MODEL variable in provider/model format (for example anthropic/claude-opus-4-8), with the matching API key. The connection logic mirrors what we detail in our guide to connecting Claude and GPT to n8n.
3. Enable the sandbox — N8N_INSTANCE_AI_SANDBOX_ENABLED=true provides the isolated environment where the assistant runs code to test what it generates, a sensible precaution when AI is producing and executing code on your own instance.

The web search feature tied to the assistant is optional and configured separately — the rest of the functionality (creating, editing, debugging) works without it. As of version 2.35, this setup has been considerably simplified compared to the manual instructions published at launch — if you're on an older version, an update is worth doing before attempting the install, following our guide to updating n8n on Docker.

Limits worth keeping in mind

Two categories of limits deserve particular attention before rolling the assistant out to sensitive workflows.

Technical limits: every suggestion stays bounded by the underlying model's context window — an already large workflow, with many nodes and past executions loaded into context, can exceed what the model can process in one pass, with less relevant suggestions as a result. For a complex workflow, it's still more reliable to split it into sub-workflows first and guide the assistant section by section, rather than asking it to generate everything at once.

Overconfidence in the generated result: a study by Vaithilingam et al., Expectation vs. Experience: Evaluating the Usability of Code Generation Tools Powered by Large Language Models (CHI 2022), found that while developers generally preferred working with a code-generation assistant, they ran into real difficulty understanding, editing, and debugging what the tool produced — a gap between expectation and actual experience. More concerning still, a study by Perry, Srivastava, Kumar, and Boneh, Do Users Write More Insecure Code with AI Assistants? (ACM CCS 2023), found that AI-assisted users produced significantly less secure code, while wrongly judging it more reliable than without assistance. The parallel to an automatically generated n8n workflow is direct: a webhook node, a credential permission, or a database query suggested by the assistant deserves the same review as a code snippet — not automatic trust just because "the AI wrote it." Our guides on securing API credentials and securing an n8n webhook remain the reference to apply, whether the node was placed by hand or generated by the assistant.

A concrete use case: prototype first, harden after

The most effective use observed for this kind of tool isn't delivering a final workflow in one command, but speeding up the prototyping phase: describe the automation, let the assistant lay down the skeleton (trigger, AI nodes, output action), then take back the wheel for the precise business logic, error handling, and credential hardening. That's where the line with a ready-made workflow pack becomes useful to understand: the assistant produces a starting point that still needs testing under real conditions, whereas the Inbox AI Pack (€79) ships, for example, AI email triage, urgency scoring, and Gmail drafts already proven in production, without the tuning phase. For a compliance or audit need, the Compliance & Audit Pack (€149) follows the same logic — already-validated workflows rather than a prototype you have to harden yourself.

Checklist before rolling the assistant out further

• Module enabled on self-hosted (instance-ai in N8N_ENABLED_MODULES) and the right provider's API key configured in N8N_INSTANCE_AI_MODEL.
• Sandbox enabled (N8N_INSTANCE_AI_SANDBOX_ENABLED=true) before letting the assistant run test code.
• On Cloud, checking your plan's actual limits before following a suggestion from the assistant that seems stuck for no reason.
• Systematic review of nodes touching credentials, permissions, and webhooks, regardless of how confident the assistant's response sounds.
• Complex workflows split into sub-workflows before prompting, to stay within the model's useful context window.

The AI Assistant changes how fast a first draft of a workflow takes shape in n8n, not who's responsible for security and production reliability — that stays on you. Used as a prototyping accelerator followed by a human review pass, it's a genuine time saver — like any code-generation tool, it shifts the effort from writing to reviewing, rather than removing it.