Repair malformed JSON from LLMs into valid JSON.
Trailing commas, single quotes, unquoted keys, markdown fences, comments, Python literals, and output that just got cut off — mended into something json.loads accepts.

▶ Try the live playground — paste broken JSON, watch it get repaired in your browser.

You asked the model for JSON. You got:

Sure! Here's the JSON:
```json
{
  'name': 'Ada',        // single quotes + a comment
  "skills": ["py", "ml",],   // trailing comma
  "active": True,        // Python literal
  "bio": "cut off mid-str

json.loads throws on every line of that. When you call a hosted model you get a string, not logits — so you can't constrain decoding; you have to repair what came back. mend does exactly that.

pip install mend-json

Zero runtime dependencies — pure standard library.

import mend

mend.loads("{'ok': True, 'items': [1, 2, 3")   # tolerant json.loads
# {'ok': True, 'items': [1, 2, 3]}

mend.repair('```json\n{name: "Ada", age: 36,}\n```')   # -> valid JSON string
# '{"name": "Ada", "age": 36}'

From the shell:

cat llm_output.txt | mend --indent 2

mend is a forgiving recursive-descent scanner that never throws on input. It strips fences and comments, jumps to the first real value, and parses leniently — tolerating stray commas, either quote style, and bare keywords. It keeps a stack of open containers and, at end-of-input, auto-closes whatever is still open (that's the trick that recovers truncated/streaming output). The result is a Python object, which repair() re-serializes as guaranteed- valid JSON.

16 malformed samples covering every defect above:

~122,000 repairs/sec, pure Python. The test suite (49 cases) asserts that every repair is parseable by the real json module and that valid JSON is returned unchanged.

python benchmarks/bench.py        # reproduce

• mend is best-effort recovery, not a validator. It assumes the input is meant to be JSON and makes it parseable; it won't tell you the model got the schema wrong (pair it with a schema check, or prevent the problem upstream with constrained decoding).
• Ambiguous cases resolve pragmatically: an unrecognized bare word becomes a string ({"status": active} → {"status": "active"}); a value buried in pure prose with no {/[ raises JSONRepairError rather than guessing.
• Numbers, strings, and nesting are preserved exactly; round-tripping valid JSON returns the same object.

make install   # pip install -e ".[dev]"
make test      # 49 tests
make lint      # ruff
make bench     # regenerate benchmarks/results.md

MIT © Hritik Datta