How to Connect the Claude API to n8n

The Claude API powers the two AI steps in every library system: writing the personalized first line of each email and classifying replies into positive, objection, not-now, or unsubscribe. Both run on Claude Haiku 4.5, the small fast model, through n8n’s native Anthropic integration.

What you need

• An Anthropic Console account at console.anthropic.com.
• A payment method. The API is prepaid, so you buy a credit balance before any call succeeds.
• A running n8n instance.
• About 10 minutes.

Get the API key

1. Sign in to the Anthropic Console.
2. Open the API Keys page from the settings menu.
3. Click Create Key, name it “n8n lead gen”, and select your workspace.
4. Copy the key now. The console shows the full value once and only a masked version afterward.

Then load credits: open the Billing page and add a starting balance. $5 is enough for months of this workload. Skipping this step is the classic trap, because the key itself looks valid and the error only appears at request time.

Create the credential in n8n

1. Credentials, then Add credential.
2. Search for “Anthropic” and select the Anthropic API credential.
3. Paste the key. Leave the base URL field at its default.
4. Name it “Anthropic LGL” so imported workflows connect on their own.
5. Save.

Test it

The reply handler uses an LLM Chain with an Anthropic chat model attached, so mirror that shape.

1. In a scratch workflow, add a Basic LLM Chain node.
2. Attach an Anthropic Chat Model as its language model and pick Claude Haiku 4.5 from the model list. The list loading at all proves the credential authenticates.
3. Set the prompt to something trivial like “Reply with the word ok” and execute.

A text response confirms end-to-end billing and access. An error about credit balance means the prepaid step above was skipped. A 401 means the key was pasted badly or has been disabled in the console.

Cost and one gotcha

The system cost tables put this line item at $5 to $7 per month, and that holds because both jobs use short prompts on the cheapest model. The gotcha is model selection drift. If you swap in a larger model to make first lines “better”, spend rises roughly tenfold for improvement your prospects will never notice in two sentences. Keep Haiku for volume work and reserve bigger models for anything a human reads at length.

Where the library uses the Claude API

The outreach workflow generates one custom opening line per prospect from its trigger data, and the reply handler classifies every inbound response so positive replies get a booking link within minutes. Some systems also use it to summarize scraped pages into the facts the email references.