Question 1

What can this tool fix?

Accepted Answer

Trailing commas, single quotes, curly quotes pasted from a document, unquoted object keys and unrecognized bare-word values, line and block comments, missing commas between values, Python spellings like True and None, and numbers written in ways JSON does not allow such as a leading plus or a bare .5. It also replaces NaN, Infinity and undefined with null, and closes a document that was cut off partway through, such as an unterminated string or an object left open at the end. What it will not do is guess at structure it never saw: a bracket missing from the middle of an otherwise complete document is reported rather than inferred.

Question 2

Is repairing safe?

Accepted Answer

For punctuation, yes. Removing a trailing comma or quoting a key changes how the document is written, not what it means. Three repairs are different: NaN, Infinity and undefined have no JSON equivalent, so they become null and your data genuinely changes. Those are listed separately in the changes so you can decide whether that is acceptable. Everything else preserves your values exactly, including long integers and the order of keys.

Question 3

Why does it say the result is still invalid?

Accepted Answer

Because it checks its own work. After repairing, the result is run through the same strict validator the rest of the site uses, and if it still does not parse you are told where. That happens with genuinely broken input, most often an unclosed bracket or a truncated file. Guessing at the missing structure would be more likely to produce something wrong than something useful, so the tool stops and shows you the error instead.

Question 4

What does the confidence badge mean?

Accepted Answer

How sure the fix is to be what you meant, not how sure the syntax problem was. High confidence covers a delimiter swapped or a comment removed, a one reading re-encoding with nothing to guess at. Medium confidence covers a character that had to be invented, such as a missing comma, or an unquoted key, which is sometimes a clean identifier and sometimes an unrecognized word quoted as a guess. Low confidence is reserved for NaN, Infinity and undefined becoming null, the one repair that changes your data rather than only its spelling, shown with its own "Changes your data" badge as well.

Question 5

What does checking against a schema do?

Accepted Answer

It runs the repaired document, not the original, through the same validator the JSON Schema Validator tool uses, and lists anything that does not match. This does not change how anything is repaired: the schema only checks the result afterward, most usefully for a lossy repair like NaN becoming null, where the syntax fix was unambiguous but whether null is actually the right value for that field is a separate question a schema can answer. Both OpenAPI and plain JSON Schema work, since an OpenAPI schema object is JSON Schema with, at most, a small set of well documented differences.

Question 6

What does "Extract from surrounding text" do?

Accepted Answer

It looks for the JSON value itself before repairing anything, for the common case of pasting a whole reply rather than only the part that matters. A Markdown code fence is tried first, since it is the strongest signal of where the payload actually starts and ends, and the common shape a pasted LLM response has. With no fence, it falls back to the first opening brace or bracket anywhere in the text and matches it to its balanced close, skipping over anything inside a string so a stray brace in a quoted sentence does not end the value early. That fallback is a heuristic, not a guarantee: a brace inside an ordinary, unquoted sentence before the real JSON could be mistaken for the start of the value, which is exactly why this is off by default rather than applied to everything you paste. Turn it on when you know the text has more than JSON in it.

Question 7

What happens with a response cut off by a token limit?

Accepted Answer

It gets closed rather than rejected. If the document ends inside a string, the closing quote is added, and a lone backslash left dangling by a cut off escape sequence is dropped rather than copied through. If it ends inside an object or array, every level still open is closed in turn, innermost first. A key that was read but never given a value gets null, marked as a change to your data, the same way NaN or undefined are. A position that was simply never reached, such as the next element of an array that stops right after a comma, gets nothing invented for it: the dangling comma is removed and the container is closed, rather than guessing at a value that was never there to guess. What this will not do is complete a cut off word like tru into true, since a handful of letters gives too little to be sure what was meant.

JSON Repair

Keyboard shortcuts

Local data

Recently opened files

Share via link

Open the shared link in

Input

Load from a URL

Output

Schema (optional)

Load from a URL

What was changed

About this tool

How to use

Keep reading