Repair LLM JSON Output: Handling Almost-JSON from AI

LLMs frequently return JSON wrapped in ```json fences, with single quotes, Python literals, or unquoted keys. Repair the syntax, validate strictly, then derive TypeScript types from the cleaned output.

When you land here

An LLM returned JSON that JSON.parse rejects. The most common offenders are markdown code fences wrapping the document, single quotes, unquoted keys, Python literals like True/None, and stringified objects. Repair the syntax, then validate strictly before you hand the result to your app.

Repair almost-JSON from LLMsValidate the cleaned JSON strictlyOpen the cleaned JSON as a treeGenerate TypeScript interfaces from the JSON

The repair workflow for LLM output

Run repair first — it tolerates fences, single quotes, unquoted keys, Python literals, and trailing commas. Validate second to confirm the cleaned text round-trips. Use the cleaned tree to derive types or to feed your app.

Repair LLM JSON output (guide)Fix unquoted keys in JSONFix single quotes in JSONEscape JSON as a string literalRepair broken JSON in JavaScriptFix JSON online (full walkthrough)

Once it parses, derive types

After the LLM's output validates, derive TypeScript interfaces or a JSON Schema so the next call has a contract to validate against. That's the difference between repairing once and repairing forever.

JSON to TypeScript interfacesJSON to JavaScript object converter (guide)Single quotes vs JavaScript object literalBad escaped character in JSONUnexpected end of JSON inputJSON.parse source-access: lossless numbers and safer LLM output

What's coming in standards

Two upcoming JavaScript proposals matter for LLM JSON workflows: JSON.parse source-access (read errors with the original text alongside the value) and JSON.parseImmutable. They reduce the amount of bespoke repair code that lives in every app.

JSON.parse source-access (Baseline 2025)Deep-dive: JSON.parse source-access in practiceJSON.parseImmutable (Stage 2)Online JSON formatter data leak — don't paste secrets

Recommended path

Paste the LLM output, clean it, type it.

  • 1. Tool: / — paste the AI output, click Repair & Format.
  • 2. Guide: /guides/repair-llm-json-output — handles fences, quotes, Python literals.
  • 3. Blog: /blog/repair-broken-json-in-javascript — the same workflow in code.
  • 4. Reference: /news/json-parse-source-access-baseline-2025 — what standard support is arriving.
Open JSON FixCompare JSON changes

JSON repair guides

Topic hubs

  1. JSON Parse Errors: Read the Message, Jump to the Fix
  2. Fix Invalid JSON: From 'What's Wrong' to a Clean File
  3. JSON Formatter, Validator, Viewer: Pick the Right Tool
  4. Privacy: JSON Tools That Don't Leave Your Browser
  5. JSON Interop: YAML, CSV, XML, JWT, Schema

Specific guides

  1. How to Decode Base64 Strings (and JWT Payloads)
  2. URL Encoding: Percent-Encode Query Parameters and Paths
  3. Convert YAML to JSON (and Avoid Indentation Errors)
  4. Convert JSON to CSV: Flatten an Array of Objects
  5. Convert JSON to XML: Root Elements, Attributes, and Arrays
  6. Escape JSON as a String Literal (and Decode Double-Encoded JSON)
  7. Fix Trailing Comma in JSON
  8. Fix Single Quotes in JSON
  9. Fix Unquoted Keys in JSON
  10. Repair LLM JSON Output
  11. Fix JSON Parse Error: Expected Property Name
  12. JSON vs JS Object Literal: The Key Differences
  13. Validate JSON Before API Requests
  14. JSON Formatter vs JSON Repair
  15. Fix JSON Unexpected Token Errors
  16. JSON to JavaScript Object Converter