How to fix “unexpected token” error when parsing JSON from Claude in n8n?

The direct fix is: stop trying to JSON-parse Claude’s raw output directly. Claude often returns normal text (with line breaks, quotes, explanations, or Markdown) mixed with your JSON. You must either force Claude to return valid JSON only, or clean the output before running JSON.parse / using the “Convert to JSON” node. In n8n, the most reliable fix is to wrap Claude’s output with a JSON schema prompt and then use a small Function node to extract the JSON block safely before parsing it.

 

Why the error happens

 

In n8n, the JSON parse error “Unexpected token … in JSON at position …” nearly always means that the text you’re trying to parse is not pure JSON. Large language models like Claude often add:

• Extra words like “Here is your JSON:”
• Code fences with three backticks
• Trailing commas or unescaped quotes
• Markdown formatting
• Multiple JSON objects in one response

n8n nodes such as Convert to JSON or Function → JSON.parse() expect a clean JSON object or array. Anything else throws the “unexpected token” error.

 

How to fix it (production‑safe approach)

 

There are two reliable ways to fix this in a real n8n workflow.

• 1. Make Claude output strict JSON only Add this to your system prompt or user instruction:
  “Return only valid JSON. No explanations, no surrounding quotes, no Markdown, no comments. Respond with a single JSON object.”
  This usually prevents dirty output, which is the ideal fix because it avoids error handling downstream.
• 2. Clean the response before parsing When Claude still adds text, use a small Function node to extract the JSON block. This is extremely common in production.

 

Safe extraction Function node

 

Place a Function node right after Claude’s output. Use this code:

const raw = $json.text || "";   // adjust key if Claude output is in a different field

// Find the first and last curly braces
const start = raw.indexOf("{");
const end = raw.lastIndexOf("}");

// If nothing looks like JSON, stop early
if (start === -1 || end === -1) {
  throw new Error("No JSON object found in Claude output");
}

// Extract substring
const jsonString = raw.substring(start, end + 1);

// Parse it safely
let parsed;
try {
  parsed = JSON.parse(jsonString);
} catch (err) {
  throw new Error("Failed to JSON.parse cleaned Claude output: " + err.message);
}

return [{ json: parsed }];

Why this works: It ignores anything before or after the JSON block (Markdown, comments, backticks, casual text) and isolates only the real JSON object. This is the single most robust method when using LLMs inside n8n.

 

Other important production notes

 

• Claude can output Unicode smart quotes which break JSON. The extract-and-parse method removes most of the risk.
• Don’t use the “Convert to JSON” node on raw LLM text. These nodes assume clean data; use them only after extraction or after a reliable structured response.
• Claude sometimes returns arrays instead of objects. If you want an object, specify it strictly in the prompt.
• Add an Error Workflow in production so you get notified instead of silently failing.

 

The short version you can tell a junior dev

 

The “unexpected token” error happens because Claude’s reply is not pure JSON even though it looks like it. Clean the output first or force Claude to output strict JSON. The most reliable fix is a Function node that extracts the JSON between the first “{” and the last “}” and then parses it.