Repair broken JSON. Review every change.

No. Scanning, repair, verification, schema validation and every export happen inside this browser tab, and no request anywhere in the tool carries the document you pasted. Opening a file reads it from your disk with the File API. The payload never appears in a URL, an analytics event or an error report.

No. It is LLM-aware, not LLM-powered. It understands the shapes model output arrives in — code fences, explanatory prose, Python literals, stringified tool-call arguments, end-of-stream truncation — because those are what break in practice. The repair itself is a deterministic rule engine running locally. No model is called, in the repair path or anywhere else.

Yes, and that is what the LLM output profile is for. There are no per-provider parsers, deliberately: the malformed patterns matter and the provider name does not. A code fence is a code fence whoever wrote it, and a parser keyed to a brand would be marketing rather than behaviour.

Formatting re-prints a document that is already valid; it changes presentation and cannot fix a broken document. Repair rewrites malformed source into valid JSON and lists every change. Validation answers whether a document is legal JSON and where it stops being legal. Schema validation asks whether already-valid JSON matches an expected structure. They are four different questions.

It can make truncated JSON parse, which is not the same as recovering it. Unclosed strings and containers are closed and the result says the structure was closed, not that the data is complete. Whatever was cut off is genuinely absent from the input, and no value is ever invented to fill the gap.

It is not applied. Ambiguous cases — undefined, NaN, Infinity, a bare identifier, a constructor call, an ellipsis placeholder — are raised as decisions with the options spelled out, and the document stays marked invalid until you choose one. The choice is recorded in the repair report.

They are reported, never discarded silently. The JSON grammar allows a repeated key and JSON.parse keeps only the last one without saying so, which means the parsed value is not what you can read. Five policies are offered — keep both and warn, keep first, keep last, rename the duplicates, or collect the values into an array — and none is selected for you.

Yes, in an optional panel that runs only after the repaired document parses. Syntax problems and schema problems are listed separately, with JSON-Pointer-style paths. The common draft-07 keywords are evaluated; $ref, allOf, anyOf, oneOf, not and the conditionals are named as unevaluated rather than quietly passed, and remote $ref documents are never fetched.

As input to be normalised, not as valid output. The JSON5 / JSONC profile removes comments, rewrites hexadecimal numbers, leading pluses and leading or trailing decimal points, folds line continuations and converts backtick strings. The result is strict JSON — the tolerant syntax is gone, not preserved.

The common cases, yes. Single quotes become double quotes, True and False become true and false, and None becomes null. Python values with no JSON equivalent — a tuple, a set, a datetime, a NaN — are not silently converted; NaN and Infinity are raised as decisions, and anything the scanner cannot classify is left for you.

Yes. When every non-empty line parses on its own, the records are collected into a single array and the count is stated. No document is merged into another, and a multi-line document that merely looks like NDJSON is left alone, because wrapping it would destroy it.

In the text, yes; in a JavaScript consumer, no. JSON numbers have no precision limit, but JavaScript parses them as IEEE-754 doubles, so an integer beyond 2^53 loses exactness on parse. The repair never rewrites the digits, and any number outside the safe range is flagged so the loss is visible rather than silent.

Yes. Every change can be rejected, and rejecting one re-runs the whole repair from the original with that change suppressed rather than undoing an edit on a working copy. If the remaining selection no longer parses, the interface says so and the copy-as-success action is disabled — a draft can still be copied, with a warning.

No, and a tool that claimed otherwise would be lying. It repairs common deterministic issues and identifies ambiguous damage for review. A document whose content is genuinely missing, whose structure is unrecoverable, or whose intent cannot be read from the text is reported as partially repaired with the reason, rather than being forced into something that parses.

The working document, the selected profile and your review decisions are kept in this browser's local storage so a refresh does not lose them. Nothing is synced to another device or to Motifuse. Private session turns local saving off for the current session, and Clear local data removes what is already stored.

Use JSON Formatter to format, validate, explore, query or convert a document, and for a quick repair when you just want it fixed. Use JSON Repair Pro when the document is badly broken, came from a model, or when you need to see and defend every change. Both run the same repair engine, so they cannot disagree, and each page has an action that hands the current document to the other.