Quick answer
Put Claude in an n8n node when a hard rule cannot classify, draft, or read unstructured junk — and give that node a threshold and a path a human will actually read. Keep the rest of the flow as rules: trigger, write-back, Slack, stop. An agent that plans its own path is a different product. We default to this shape on AI automation. We do not add a model because the workflow looked boring.
Key takeaways
- The model is a step, not the workflow. n8n still owns the trigger, the branches, and who gets told.
- Claude earns its seat on judgment: messy text, a draft, a class you cannot if/else. A hard rule stays a rule.
- If you cannot see what it decided, you do not have automation. You have a coin flip with an API bill.
- A threshold and a human path are part of the node. Without them you will ship a confident mistake.
- An agent that picks its own tools is not “n8n plus Claude.” It is a ceiling you should name on purpose.
What “a step” means
A step has a job, an input, an output shape, and a failure. In n8n that is a node: the payload arrives, Claude returns JSON you defined, the next node writes the CRM or posts Slack. The model does not choose the next tool. The flow does. That is why we use n8n as the runner and Claude as the judgment — spelled out as n8n automation and the automation offer.
A chat widget in the corner of the site is not this. A Zapier “AI by Zapier” bolt-on can call a model. The gap is the log, the fallback, and whether you can host the runner. That compare is n8n vs Zapier vs Make. This page is the node itself.
| Rule | Claude step | Agent | |
|---|---|---|---|
| Who decides the path | You, in the canvas | You. The model fills a field | The model, until you stop it |
| Good for | Known shapes, exact matches | Messy text, a draft, a class | A goal with several tools and a budget |
| When it fails | You wrote the miss | Threshold, retry, human | It invents a path you never drew |
| Cost to debug | Read the node | Read the prompt and the log | Read a story |
| Wrong reason | "AI is overkill" | "We should use AI" | "Agents are the future" |
When Claude earns the node
A form reply, an inbound email, a PDF. You cannot if/else “kind of a customer.” The model returns a label and a confidence.
A first pass the human edits. Sending without a path to stop is how you email the wrong invoice.
A site, a thread, a contract clause. Extract fields you named. Do not ask it to “handle the lead.”
If the next step is “plan is Enterprise, owner is Sam,” write the rule. Models are expensive to run and expensive to explain.
The test: can you write the output as a small schema? Label, score, three bullets, a yes/no. If the answer is “it should just figure it out,” you do not have a step. You have a wish. Prompt engineering here is a spec: role, shape, bad examples. Not a witty paragraph.
The threshold and the human path
Every Claude node we ship has a number and a door. Below the number, the flow does not pretend it knew. It flags, it queues, it Slack-mentions the person who still owns the miss. Above the number, it writes the field and moves. You will tune the number after a week of real runs. That is expected. Shipping without the door is not.
- 01Name the field the model is allowed to fill
One job. Segment, draft, extract. Not “process the lead.”
- 02Write the fail
Low confidence, empty input, provider down. Each one has a branch you drew.
- 03Log the prompt, the input, the output
If you cannot replay a bad run, you cannot fix the prompt. Fashionable nodes hide this.
- 04Put a human on anything that sends, spends, or deletes
Classification into a CRM field can go. An outbound email cannot, not on week one.
“If you cannot point at the judgment, you do not need a model. If you can, n8n is the runner we use — and the model is still just a step.”
Chris NutbeenNuttifox
When an agent is the wrong noun
An agent keeps a goal, picks tools, and continues. That is useful when the path cannot be drawn in advance — research, a messy internal hunt. It is a poor default inside ops. Most “agentic workflow” pitches we see are a classify-then-route flow wearing a trench coat. Draw the path. Put Claude on the one judgment. Keep the rest boring.
If a spreadsheet formula would do it, skip the model. If nobody will read the Slack when it is unsure, skip the model — you will not operate the fallback. If the data cannot leave the building and you will not self-host, skip the cloud call. If you wanted a chatbot in the product, that is AI development, not an n8n node.
n8n Cloud Starter is €20 a month billed yearly for 2,500 executions, on the list we checked in Sep 2026. Claude is a usage bill on top. Studio time sits on the contact page. A week that adds one honest node to a mapped process is a good week. A week that sprinkles models on every branch is how you get a flow nobody will turn back on after the first wrong send.
FAQ
Can n8n use Claude?
Yes. A node calls the API, you define the output, the next node does something deterministic. That is the whole trick.Is this an AI agent?
Usually no. An agent picks its own path. A Claude step fills a field you named. Most automations should be the second. Agents are a product decision, not a node setting.Why Claude and not ChatGPT?
Claude is our default for judgment and instruction-following. If you already run another model, we use that. The boundary matters more than the badge.Will it send email on its own?
Not on a flow we ship in week one. Drafts go to a human or a queue. Sending is a later privilege you earn with a log you trust.Do we have to self-host n8n to use Claude?
No. Cloud is fine if the payload is allowed to leave. Self-host when procurement or the data says so. The step looks the same.Can you add this to our existing zaps?
We can rebuild the flow that earns, with the judgment as a node. We will not drop a model into a zap we have not mapped. Read the n8n compare if you are still on Zapier.How much does a Claude step in n8n cost?
n8n Cloud Starter is €20 a month billed yearly for 2,500 executions, on the Sep 2026 list. The model is extra, billed by tokens. The week that wires the node is the studio. Ours is on the contact page.Have something to build? Let’s ship it.
Book a call