Long ChatGPT conversations that stop lagging — and everything you can do with them afterwards.
Trim heavy history, search in place, jump by question, pin replies,
and export to Word / PDF / Markdown / HTML / TXT / JSON — with images, from your own device.
Screenshots · Languages · 简体中文 · Changelog · Contributing · Privacy
The poster above is the original launch artwork: its Folders / 文件夹 tile and its
v4.0badge are stale — the folder organizer was retired before this release, and the extension version is1.0.0. Everything shown under Screenshots is rendered from the real UI.
Free forever. No account, no subscription, no ads, no upsell screen, no telemetry. Everything runs in your browser and nothing you write ever leaves the device.
Qelvynox was called TurboGPT previously. Existing installs carry their settings and bookmarks forward automatically on the first load.
| Navigate by question | Search in place |
|---|---|
 |
 |
| Pinned replies | Export |
|---|---|
 |
 |
| Control panel (light) | Control panel (dark) |
|---|---|
 |
 |
The interface ships in Chinese and English, and you switch it inside the product: Settings → General → 界面语言 / Interface language, or the segmented control at the top of the toolbar popup. It applies immediately — no reload — and is remembered.
- Chinese is the default, whatever your browser language is. The choice is a
stored setting, never a guess from
navigator.language. - The switch covers Qelvynox's own panels, tooltips, notifications and the labels in exported files. Your conversation text is never translated.
- The store listing and Chrome's own extension UI are localised separately through
_locales/zh_CNand_locales/en.
Adding a third language means one more table in src/shared/i18n.js, one more entry
under _locales/, and nothing else — no string lives anywhere else in the code.
ChatGPT renders every turn it has loaded, which is why a 500-message chat turns into a slideshow. Qelvynox keeps only the most recent turns in the layout and leaves the rest in the page's memory — without deleting anything ChatGPT owns.
- Older turn wrappers are hidden and the virtualizer rail is recalculated, so
React's fibre tree stays valid. Physically removing nodes is what made ChatGPT
throw
NotFoundErrorand show "Oh no! An error occurred". - Three loading modes: recent turns only, ChatGPT default, load everything.
- Scroll up and older turns reveal themselves in batches, with the reading position anchored to the pixel. No button, no jump.
- Background tabs do nothing at all until you come back.
Type a keyword and every occurrence in the conversation is painted in place;
the first one is centred, Enter / Shift+Enter walk the rest.
- Painted with the browser's Custom Highlight API, so ChatGPT's DOM is never rewritten and React can keep re-rendering underneath it.
- The scope is every turn the page has loaded — including the ones the booster folded away, which are revealed on the way to a match.
- A reply that arrives while the bar is open is searched too: the bar repaints as the answer streams in, throttled so it never fights the page.
Hover the dock's outline button and a headerless list of your questions slides
in; click to pin it. The question nearest the middle of your viewport is
highlighted and kept in sight while you scroll, so the panel works as a position
map rather than a jump list. Split views are handled too: the list only ever
describes the conversation you are looking at, never the sum of both. Clicking a row scrolls straight to that turn with a
glowing highlight, unhiding it first if the booster had folded it away.
Rows are also tab-stops: Tab walks the list, Enter jumps.
A star button on every assistant reply. Pinned replies live on their own dock
button with a 本对话 / 全部 switch — a bookmark from another chat opens that
chat, one from this chat jumps to it and reveals it first.
Asking OpenAI for a conversation again is the request that gets rate-limited, refused or cut short on a long chat. The local archive removes that request:
- Captured from the payload ChatGPT already loads when a chat opens — no extra request, no extra latency.
- Kept current as you chat: each message is saved once it finishes streaming, and one throttled sync pulls the newest turns with their real ids.
- Older history is filled in the background, one page at a time, resuming from the oldest saved page on the next visit.
- Images are stored as bytes while their signed URLs are still valid.
- Two instant exports read only from your device: a self-contained HTML file with every image embedded (opens offline, forever) and Markdown with a marker where each image was.
- Temporary chats are never archived.
Delete archiveis one click away.
Tick exactly which messages you want, then export to:
| Format | Notes |
|---|---|
| Word (.docx) | Genuine OOXML built from scratch — minimal ZIP writer, CRC32, no library. Code blocks keep monospace and shading. |
| Print view, formatted markdown rather than a flattened page. | |
| Markdown / TXT / JSON | Structure survives: fenced code keeps its language, links keep their URLs, lists and tables keep their shape. |
| HTML | From the local archive, images embedded, right-to-left text laid out correctly. |
For a chat that hit ChatGPT's length limit and will not take another message:
- Full conversation (.md) walks every turn from the conversation API — including the ones trimmed for speed — without loading them into the page.
- Continuation prompt copies a ready-to-paste block with the last N turns plus an instruction to continue rather than restart.
Start a new chat, attach the .md, paste the prompt, carry on. If the walk
cannot finish, the export says exactly how far it got and why; a partial archive
is never presented as complete.
Qelvynox does not send conversation data to the developer. There is no analytics, no crash reporting, no update ping, no remote feature flag, and no server of any kind in the codebase. Your prompts, replies and images are processed in the browser and may be stored on your device. History and image requests go to OpenAI services using your current session. See docs/privacy.md for the policy text and docs/store-listing.md for the exact store disclosures.
| Permission | Why it is needed |
|---|---|
Page access on chatgpt.com / chat.openai.com |
That is the whole product: the toolkit renders inside the conversation. |
storage, unlimitedStorage |
Settings and the local archive are stored on the device. |
Host access to *.oaiusercontent.com |
Fetching image bytes for the archive. HTTPS only, ChatGPT/OpenAI hosts only, requests only accepted from this extension's own scripts. |
No tabs, no activeTab, no browsing-history access, no broad host permissions,
no remote code — every line ships inside the package.
From the Chrome Web Store (recommended): search for Qelvynox and click Add to Chrome. Works in Chrome, Edge, Brave, Opera and Vivaldi.
From source (no build step, no dependencies):
git clone https://github.com/WeFode/Qelvynox.git # mainland China: git clone https://gitee.com/lin_30_0/qelvynox.git- Open
chrome://extensions, enable Developer mode. - Click Load unpacked and select the repository folder.
- Open or refresh ChatGPT. The dock appears on the right.
| Key | Action |
|---|---|
Alt + F |
Search this conversation |
Enter / Shift + Enter |
Next / previous match |
Tab + Enter |
Walk and jump the navigator list without a mouse |
Esc |
Close search, navigator, bookmarks, export and settings |
| Path | Role |
|---|---|
src/shared/i18n.js |
The dictionary: one parallel table per language, plus the switch, placeholder/plural handling and the popup's data-i18n pass. Loaded first among the content scripts. |
src/page/mainWorld.js |
MAIN-world fetch interceptor that trims conversation payloads before React renders them, and hands the original payload to the archive. |
src/content/index.js |
The in-page suite: dock, navigator, bookmarks, search, export, settings panel — driven by one debounced MutationObserver. |
src/content/archive.js |
The local archive: message identity, merge rules, safe HTML/Markdown builders, storage adapter. |
src/content/settings.css |
The settings panel's design tokens and components, namespaced under .qgst-* so nothing leaks into ChatGPT. |
src/popup/ |
A compact launcher: quick switches plus the door into the in-page panel. |
src/background/background.js |
Settings defaults and migration, and the image fetcher for the archive. |
node tests/run.mjs # 11 suites, 543 assertions, zero dependenciesEach suite loads the real extension source into node:vm with stubbed
browser globals and drives it with a scripted network and DOM — payload
trimming, backwards pagination, turn counting, hydration, stats plumbing,
temporary chats, archive merge and safe HTML, the export pipeline, the settings
panel, live-surface refresh, the language dictionary and the release manifest.
node dev/serve.mjs # http://127.0.0.1:8123/c/demo-chat
node dev/shots.mjs # regenerate the screenshots above from the real UIdev/ never ships with the extension.
| Host | URL | Role |
|---|---|---|
| GitHub | https://github.com/WeFode/Qelvynox | Primary — issues and pull requests go here. |
| Gitee | https://gitee.com/lin_30_0/qelvynox | Mirror, for a faster clone in mainland China. |
Both remotes are configured in the maintainer's working copy as origin (GitHub)
and gitee (Gitee), on the main branch:
git remote add origin https://github.com/WeFode/Qelvynox.git
git remote add gitee https://gitee.com/lin_30_0/qelvynox.git
git push -u origin main # first push
git push -u gitee main # keep the mirror in stepdocs/store-listing.md is the whole checklist: the artefacts and their exact sizes,
the English and 简体中文 listing text, the per-permission justifications, the
data-use answers and the items to verify before clicking Publish. The short version:
node tests/run.mjs # must be green
node dev/package.mjs # -> qelvynox-<version>.zip (no dev/, tests/, docs/, screenshots or posters)Then upload that zip in the Chrome Web Store dashboard, fill both item languages from
the doc, and keep the privacy-policy URL public. The store wants a page that loads
without a login, so publish docs/privacy.md through GitHub Pages
(https://<user>.github.io/Qelvynox/...) rather than pasting a blob or
raw.githubusercontent.com link.
MIT — free for personal and commercial use.
Copyright (c) 2026 WeFode (Qelvynox) and contributors. Qelvynox is an independent, unofficial tool and is not affiliated with or endorsed by OpenAI.