v0.2.3

Added

• Firefox and Edge support. Each release now has a package for Chrome, Edge and Firefox. Firefox needs version 140 or later and gets an .xpi signed by Mozilla for self-distribution, so it installs for good rather than as a temporary add-on; its install prompt declares that website content is sent elsewhere — the text you translate, going to the service you chose. The Edge package is the same build as Chrome's, which Brave and Arc also use. The update check offers Firefox the .xpi and the other browsers the Chromium package.
• Several models per provider. Each provider keeps a list of models, and selection cards, full pages and subtitles each choose a provider and a model — a strong model for selections and a cheap one for whole pages under the same key, say. The card's provider menu lists every model as well, and each model's answers are cached apart. Existing settings carry over: a provider's single model becomes a one-item list, and each mode keeps the provider and model it was using.
• The endpoint's own model list. Once a provider's key works, the models its endpoint offers are listed beside yours, with a search box; click one to add or remove it. A provider being set up switches itself on as soon as its key works and a model is chosen — one you have switched off stays off.
• Vendor icons for models, from LobeHub's own model table: about a hundred vendors, matched by model name on any provider — OpenRouter, SiliconFlow, a local runtime or a vendor's own API.

Changed

• Nothing is switched on or chosen in a new installation. Microsoft and Google start switched off like every other provider, and no mode has an engine, so no text leaves the browser until you choose where it goes. For a key-free start, switch one on under Providers and choose it for each mode under Translation engines. Existing installations keep their settings.
• A mode whose provider or model goes away is not configured, with no fallback. If no engine is chosen for a mode, the provider it uses is switched off — by hand, or because its last model was removed, which now switches it off — or its model is removed or no longer offered by the provider, the engine picker shows Not configured, with one note below the pickers explaining it, and the card, the page and the subtitles report it instead of translating. Nothing is sent anywhere until you choose again; switching the provider back on does not restore the old choice. Switched-off providers no longer appear among the choices.
• Redesigned settings. General shows one live preview — the card at its chosen size over a page, with the trigger icon or your shortcut — above plain rows, and the shortcut fields appear only in hotkey mode. Data uses segmented steps for cache age and tiles for export and import. A provider opens into separate cards for connection, models and reasoning.
• Providers are grouped into free translation services, cloud APIs (bring your own key) and local or custom endpoints. The notice about the free services moved there, shown on request, and no longer sits under the engine pickers.
• The popup leads with the current page: the site's name, a larger Translate full page switch and a line saying which state the page is in. The engines sit in a card of their own; in hotkey mode the footer shows both shortcuts.
• Shorter READMEs in all 14 languages, with a badge for each supported browser and install steps per browser. Localized READMEs live in readmes/<locale>.md; the detail moved to English documentation in readmes/develop/.
• Safer dependencies. Build and test tooling is patched — pnpm audit went from 92 findings to 2, none of them in code the extension ships — ESLint moves to 10 and WXT to 0.21, and installs take only releases that have been public for at least a week. CI now builds with pnpm 11 and Node 22, as the new install settings require.

Fixed

• A switched-off provider was still used. Turning a provider off only removed it from the pickers: modes already routed to it — Microsoft included — kept sending text to it. A switched-off provider is now never used.
• A change made in the popup could be undone by the settings page: with both open, the next change made on the settings page wrote back its older copy of everything else. Each page now writes only what it changes and shows changes made on the other.
• The Kimi icon was invisible on a light background; only its blue dot showed.
• The trigger-mode label no longer carries a "(default)" note that ran past the edge of its control.

v0.2.2

Added

• Complete READMEs in all 14 interface languages, with a centered project header, language navigation and links to installation, configuration and project activity. Localized files use README_<locale>.md; the Simplified Chinese entry is now README_zh-CN.md instead of README_CN.md.
• Custom base prompts with a template library and editor. The default is visible and read-only; create from it or from a blank template through New. Save & apply in one step, or save without applying for later. Internal dictionary and output-format rules remain managed by the extension.
• 56 translation targets, including US, UK and Australian English, and 14 interface languages. New UI locales: Brazilian Portuguese, Italian, Russian, Turkish, Vietnamese and Indonesian. Search target languages by native, English or localized name, code, or common English-region aliases.
• Subtitle preview and compact controls for both lines, with exact size and opacity inputs, a small color palette, HEX entry and a visible reset button.

Changed

• A compact interface with system UI fonts and five settings pages: General, Translation, Providers, Subtitles and Data. Provider setup is separate from translation policy; cache controls now live under Data.
• Popup focuses on target language, a current-page bilingual switch and compact engine choices. Trigger mode stays in settings. Version and update controls appear only in settings, using the same font as the interface, and what a check finds lands in a full-width strip below the header rather than inside it — so the header keeps one height and long localized instructions are not squeezed into a narrow column.
• Removed same-language blocking for selections, pages and subtitles so mixed-language text and regional rewrites can be requested normally.
• Generic English targets migrate to US English. Unsupported English varieties on free services silently fall back to generic English without switching providers. LLMs receive explicit regional instructions.

Fixed

• Passages no longer enter dictionary mode. Short terms retain automatic lookup, but dictionary entries must match the whole selection. Suspect structured output is validated before rendering; invalid output gets one text-only correction and is never exposed as raw protocol JSON or cached.
• Page and subtitle batches align by ID. Missing/duplicate IDs and invalid value types are rejected. A failed batch uses validated text-only fallbacks instead of displaying raw responses. Retry attempts replace earlier streamed text rather than appending to it.
• Substantial mixed-script source text can display a mixed-language label instead of being confidently identified as English by character count.
• Prompt edits use separate LLM cache entries. Page and subtitle instructions are also isolated; free-service caches are independent of custom prompts.
• Subtitle settings show valid values from the player instead of blank dropdowns when a value is absent from a short preset list.

v0.2.1

Added

• Check for a new version, from the settings header. Opposite the title: the installed version, and a button that asks GitHub whether a newer release exists. Only when you press it — nothing runs on its own, and nothing about you is sent. If there is one, it offers the build itself: one click downloads the .zip, and a note says what to do with it, because Chrome cannot install an extension it did not install in the first place.
• DashScope now offers every region it publishes, and its subscription. Beijing, Singapore, Hong Kong and US (Virginia) are listed separately, and so is Token Plan — a subscription on hosts of its own, whose sk-sp- keys the metered endpoints refuse and which refuses metered keys in turn. Frankfurt and Tokyo are reachable only at a per-workspace address, so this provider also lets you type one in. Each endpoint is a separate account with its own key and model list, and the form now says so.
• A custom or local endpoint can say which reasoning parameter it reads. Thinking control was previously only available for the vendors we had the mapping for; for a custom endpoint or a local runtime, what the server accepts is decided by software you chose, so you can now name it — enable_thinking, reasoning_effort, thinking.type and the rest — or leave it sending nothing, which is still the default.

Changed

• The selection card is one fixed size. Its height used to come from whatever was in it, capped by however much window was left below the selection — so the same word came out four times taller at the top of a page than at the bottom, and every streamed answer grew the box line by line. Now the frame never moves: the original and the translation each scroll in their own half of it, and the card goes above the selection when it will not fit below. The original is capped at three tenths of the space the two panes share — it is there to check the translation against, not to be read instead of it. A card still waiting for its first words is the one exception to all of this: it shows only that it is waiting, at the size of saying so, and the frame arrives with the first words. Copy moved up beside translate-again, which leaves both panes the same width and their scrollbars lined up.
• How big the card is, is now yours to set (General → Appearance). Six widths and six heights. Fixing the size is what keeps an arriving answer from moving anything, but how much room that answer deserves depends on the screen it is read on.

Fixed

• Copying from the card no longer pastes a black background. Chrome writes an HTML flavour alongside the plain text on a copy, and it carried the card's own background into whatever you pasted into. A copy that starts inside the card now yields the words alone.

v0.2.0

Added

• Many providers at once, and one per job. Providers are now a list you configure independently rather than a single active profile, and each of the three surfaces — the selection card, full-page translation, and video subtitles — is routed to whichever one you want. Routing lives in General settings, because which provider answers is a policy about your reading, not a property of any provider. New arrivals: Anthropic (Claude), Gemini, and opencode, whose Zen and Go plans are separate products with separate catalogues and are offered as separate endpoints. The Providers page is a searchable panel of fixed height, so a growing list never pushes the rest of the page off the screen, and every provider that is switched on reports its own latency.
• 34 target languages, chosen separately from the interface language. What you read in and what the extension speaks to you in are different questions.
• The card answers on its own terms. The credit line at its foot became a provider switcher — who answered and who could answer instead are the same question — with the target language at the other end of that line, a copy control beside the translation, and translate-again in the header. Every one of those choices lasts as long as the card and is never written back to settings: a provider tried on one paragraph is an answer about that paragraph, not a change of mind about the extension.
• The card names the language pair, detected from the selection itself so it is there before the answer is, and so it works with a model that never reports a detected language.
• Subtitles beyond YouTube. The translator now takes a site rather than assuming one. Zoom cloud recordings work — its transcript lives behind Zoom's own endpoint rather than on the video — as do Canvas recordings, which Canvas serves from an iframe of their own, and any player that ships its captions the standard way with a <track> element. The toggle joins the player's own control bar where there is one to join, and takes a corner of the picture where there is not.

Changed

• The popup lost its status strip. Naming one provider could not stand for three, and the routing section below names all of them.
• Card dropdowns open downward, so choosing a provider never covers the translation you are choosing it for.

Fixed

• Asking again now reaches past the cache, which would otherwise hand back the same words.
• The "already in your target language" notice no longer hides the language control, which is the one thing that can answer it.
• A second translation run can no longer have its chunks overwritten by the run it replaced.

v0.1.9

Added

• Translate without an API key. A new translation-engine choice (Settings → Translation, and in the popup) picks between Microsoft, Google, and your own model. The two free engines need no key and no configuration, so a fresh install translates the moment it is loaded; selection, full-page and YouTube subtitles all go through whichever engine is selected. Your own API remains the only engine that does dictionary lookups, since those depend on the model deciding what a selection is. These free options call public endpoints rather than official APIs — this project is not affiliated with Microsoft or Google, the endpoints may change or stop working without notice, and text you translate is sent to those services under their terms. See "A note on the free engines" in the README.
• Auto-generated (ASR) YouTube captions are now translated. They used to be skipped for two reasons, both fixed: they render as a rolling window that covered the injected line, and they arrive a couple of words at a time, which is not enough context to translate. The subtitles are now drawn by the extension rather than injected into YouTube's caption markup, and ASR words are regrouped into sentences — split on punctuation, on a pause, or on length — before being sent.
• A settings menu inside the YouTube player, on the extension's button in the control bar: a switch for the subtitles, and a style page with display mode (bilingual, original only, translation only), which line sits on top, backdrop opacity, and size, colour, font and weight for each line separately. The same settings are on the options page.
• The subtitle block can be dragged by the grip above it, to either half of the picture — dragged past the middle it anchors to the top and stays there when the player grows. The position is kept as a percentage of the player's height, so it means the same thing inline as in fullscreen, and the block lifts clear of the control bar while that is on screen.
• The translation card can be moved by its grip and pinned. A pinned card holds its place on screen while you scroll, survives a click elsewhere, and translates the next selection without moving.
• The card shows the source text above the translation, clamped to three lines with a chevron to open it, and names whatever produced the translation — the model for an LLM, the service otherwise.
• Each provider shows its own mark in the picker, from Lobe Icons (MIT).

Changed

• The theme system is deferred. Four palettes with uploadable variants, two token-injection points and a service worker that recoloured the toolbar icon pixel by pixel was a lot of machinery to carry through every new surface. It may come back in a form that costs less to carry; for now one palette remains, in light and dark, still following the system or your choice. Latin text is now set in JetBrains Mono (SIL OFL, bundled) and CJK in Source Han Serif — the family Google also publishes as Noto Serif CJK, SIL OFL. Like every CJK face here it is named rather than shipped: they are megabytes each. Corners are rounder throughout.
• The engine picker separates translation services from LLM providers rather than listing all three as peers, and the API fields are hidden — not just dimmed — when a service is selected.
• The provider dropdown is a real list showing each vendor's mark. A native select can only hold text.
• The settings page is one page per subject instead of everything that was not the API key piling up under "general": Translation (engine, target language, cache), General (trigger, shortcuts, appearance), Video (subtitles) and Data (import/export). Type is a notch larger throughout, section headings most of all.
• YouTube subtitles are set in the player's own caption font by default, so the translated line reads as a pair with the caption above it. Nothing is bundled — the family is named, and resolves to whatever the page and the system already provide.
• YouTube subtitles are translated around the playhead — roughly the next 30 seconds, widened when the video plays faster — instead of the whole transcript up front. Subtitles start appearing within a second or two on a long video instead of after the entire track has been processed, scrubbing lands on the new position immediately, and nothing is spent on parts of the video that are never watched.
• YouTube subtitles are positioned and sized from the player rather than from YouTube's caption elements, so theater mode, fullscreen and resizing need no special handling.

Fixed

• Dragging the subtitles could close the translation, for three separate reasons. Releasing a drag writes the new position, and the content script rebuilt every watcher on any stored change — so the drag tore the translator down mid-video. The drag also leaked a click onto the player, which pauses the video and, over an end screen, navigates away. And the overlay took YouTube's styles and event delegation, so it now lives in a shadow root.
• The translation card closed when you pressed its own pin, grip or source expander. Its outside-click check compared the event target against the shadow root, but a document listener sees that target retargeted to the host — so every press inside the card read as a press outside it.
• A subtitle line whose translation request failed used to stay on "translating…" for the rest of the video. Each cue now gets one retry and is then shown as the original alone.
• The button at the foot of the in-player style page reset only the block's position, so changing the subtitle size and colour and then pressing it appeared to do nothing. It resets every setting on the page. Its scroll track also ran into the panel's rounded corner and was clipped by it.
• Japanese and Korean had no font of their own anywhere in the extension. Korean fell through to the monospace tail of the stack, and Japanese Han characters were drawn with Chinese shapes — 直 and 骨 are not the same glyph in the two languages. The font stack now follows the language of the text: the interface language for the extension's own surfaces, the target language for the on-video subtitles.

Note on upgrading

• The manifest now requests access to edge.microsoft.com and translate-pa.googleapis.com, the two free translation endpoints. Chrome asks you to approve the new permission when the extension updates.
• Cached translations from earlier versions are ignored (the cache key now includes which engine produced the entry, so results from different engines can't collide). They expire on the usual schedule.

Install: download the .zip below, unzip, then load it at chrome://extensions → Developer mode → Load unpacked.

v0.1.8

Added

• Model thinking control, off by default: hybrid reasoning models (e.g. DeepSeek V4) think before every answer unless told not to, which made translations noticeably slower and billed invisible reasoning tokens. The extension now sends the provider's disable parameter by default for DeepSeek, Zhipu, Qwen (DashScope), SiliconFlow, and OpenRouter, restoring fast non-thinking replies. A "Model thinking" control (Settings → API, and in the popup for supported providers) turns thinking back on at five effort levels — Low / Medium / High / XHigh / Max — mapped to each provider's native parameter (reasoning_effort, thinking_budget, or reasoning.effort); the choice is remembered per provider. Providers without a safe parameter (OpenAI, Moonshot, Mistral, custom, local) are unaffected.
• Themes: four built-in themes — Cobalt (default, the previous look), Graphite, Sepia, and Teal — in Settings → General → Appearance. Each theme styles the popup, the settings page, and the in-page translation card, defines both light and dark variants (the light/dark mode setting keeps working independently), and carries its own typography — Cobalt keeps Geist, Graphite pairs a neutral grotesque with Consolas, Sepia is a serif with a typewriter mono, Teal a humanist sans. You can also upload your own theme as a JSON file and delete it again; the format (documented in the README) is validated strictly — all 12 color tokens required in light, unknown fields rejected. The extension's toolbar icon follows the theme too: it is tinted to the active theme's brand color using the same light/dark variant the page resolves, matching the in-page trigger icon.

Removed

• Prompt templates. The four built-in styles and custom templates are gone; translation always uses one built-in prompt, and the model judges register and topic on its own. The Prompts settings page, the template pickers, and the related storage fields were removed — existing settings (and old settings-export files) load cleanly, with the legacy fields stripped.

Install: download the .zip below, unzip, then load it at chrome://extensions → Developer mode → Load unpacked.

v0.1.7

Changed

• YouTube subtitle translation now works with manual / creator-provided caption tracks only. The translate button no longer appears on videos that have just auto-generated (ASR) captions — YouTube renders those in a rolling style that covered the translation line, so bilingual could not show. Manual-caption videos are unaffected.
• The Base URL hint in settings now matches the selected service type: cloud providers in Cloud mode, and self-hosted runtimes (LM Studio, Ollama, llama.cpp, vLLM) in Local mode (previously it always listed cloud providers, including Ollama by mistake).

Install: download the .zip below, unzip, then load it at chrome://extensions → Developer mode → Load unpacked.

v0.1.6

YouTube subtitle translation

On a YouTube video, click the translate button in the player controls to translate the video's existing captions. The translation appears as a second line beneath YouTube's native subtitle (bilingual), matching the caption font, with a "translating…" placeholder while it works.

• The line you're currently watching is translated first, and it keeps up when you seek.
• Requires the video to have captions (manual or auto-generated) with captions (CC) turned on.
• Works with your configured model — translation concurrency is provider-aware (cloud fans out across batches; local runs a shallow pipeline).
• YouTube only; it does not transcribe audio (no speech-recognition).

Load the attached browser-translate-0.1.6-chrome.zip as an unpacked/zipped extension in a Chromium browser.

v0.1.5

Added

• Full-page bilingual translation — translate a page's main content in place. The original stays, with the translation inserted under each block, rendered progressively as you scroll. Page chrome (navigation, headers, footers, sidebars) is left alone, and blocks already in your target language are skipped. Toggle it from the popup ("Translate page" / "Show original") or with the Alt+A hotkey (in Hotkey trigger mode).
• Press-to-record keyboard shortcuts — set hotkeys by pressing the keys (Settings → General).

Changed

• Selecting a long passage that is already in your target language now shows a brief "no translation needed" note instead of calling the model. Dictionary lookups of single words are unaffected.
• The popup leads with translation controls; API configuration moved below. Hotkey settings now live in Settings only, and both the selection and full-page hotkeys are active only in the "Hotkey" trigger mode.
• Cache default retention is now 7 days (was 30); still configurable.

Removed

• Translation history. The cache already avoids re-charging for repeat translations; dropping the stored history reduces what the extension persists and simplifies the app.

v0.1.4

Added

• Interface language in 8 languages — Simplified/Traditional Chinese, English, Japanese, Korean, Spanish, French, German. "Follow system" auto-detects the browser locale; the translation card follows the chosen language too.
• More cloud provider presets — Moonshot (Kimi), Zhipu GLM, Qwen (DashScope), SiliconFlow, OpenRouter, and Mistral, alongside OpenAI and DeepSeek. Providers with separate China / International services offer an endpoint picker.

Changed

• Auto dictionary/translate — the model decides whether a selection is a word/term to define or text to translate, in one streaming pass (replacing the client-side guess). The manual Translate/Define toggle and the streaming on/off setting are removed; translations always stream and are revealed progressively in the card.
• Dictionary entries now include the term's formal translation in your language, shown separately from the explanatory definition.
• The target translation language list is unchanged.