What's new in QDev.

QDev 1.0 — huge files, on-device AI, and local files that just open.

• JSONL (newline-delimited JSON) is now a first-class format: open a .jsonl

file and every record is browsable in the tree, table, and columns views.

• AI on very large files: ask and search now work on huge JSON/JSONL (hundreds

of MB to multi-GB). Answers are grounded on the *whole* file, not just the

first screenful — so fields that only appear deep in the file are found. As

always, your data never leaves the device; only field names and shapes inform

the AI.

• Exact counts and fast navigation on huge files: the status bar shows the true

record count instead of a truncated estimate, and jump-to-end/line is a quick

seek instead of a freeze.

• Large local files open from disk: opening a big local .json in the browser

now shows it in QDev instead of Chrome's raw text; a file too large to render

as a page offers a one-click drop-to-open that streams it at any size.

• Faster on big files: snappier opening, smoother keyboard navigation, and

quicker repeated queries.

• A file too large for a browser tab (over ~4 GB) now shows a clear message

instead of crashing the tab.

• On-device AI with no setup: if your browser ships Chrome's built-in model

(Gemini Nano), QDev answers entirely locally — no key, no account. Clearer

Ollama setup guidance when a local daemon blocks the browser.

• Under the hood: the viewer is now a standalone, AI-free core with optional

plugins — faster to load and fully open-source. No change to how you use it.

Two new ways to explore documents, an editor-style menu bar, and a fix that

keeps you signed in.

• New Table view: browse an array of objects as a sortable table — click a

column header to sort, double-click a row to jump to it in the tree.

• New Columns view: navigate nested data Finder-style, column by column, with

arrow-key navigation and a detail panel showing each value's type and path.

• New menu bar (File / View / Settings / Help) replaces the view tabs: open and

save files, switch views with ⌘1–⌘6, search every menu command from Help, and

check version and account info in the new About dialog.

• Fixed: the extension no longer shows you as signed out (hiding your Pro

features) a few hours after signing in — your session now renews itself in

the background, and a brief network hiccup can no longer flip you to

signed-out either.

• Fixed: asking "how many match?" on very large documents now returns an exact

count instead of silently undercounting when the result list was truncated.

AI on large documents, smarter recovery from oversized results, and an opt-in

program to help improve query accuracy.

• AI ask and search suggestions now work on large documents. Previously

anything over 64 MB showed "AI query isn't available for very large

documents" — now questions are grounded on a compact structural sample and

queries stream over the file, so even multi-hundred-MB documents are

askable. As always, your data never leaves the device — only field names

and shapes inform the AI.

• Questions whose answers are too large to display (over 25 MB) now return a

bounded preview instead of failing with an error.

• New in settings: optionally share de-identified question + schema data from

your queries to help improve jq accuracy — off by default, with billing

discounts for Pro subscribers who opt in.

• A polished loading screen while large files open.

Initial release.

QDev is a local-first JSON, YAML, and TOON viewer that runs entirely in the tab:

tree view, cmd+K search, schema inference, jq queries, and an interactive graph —

no network access for any of it. An optional AI mode turns a natural-language

question into a jq query; the generated query runs locally against your document.

• Tree, graph, and raw views for JSON/YAML/TOON of any size
• cmd+K fuzzy search across keys and values
• One-click JSON Schema inference
• Run jq queries locally (WebAssembly), plus a SQL view over arrays
• Optional AI "ask" mode: on-device (Chrome built-in, Ollama) or hosted; only a

value-masked schema and your question ever leave the device, never the values