Read the web, your PDFs and your books in your own language, without sending a word of it to anyone. No account, no API key, no cloud.
Real time, unedited, on the default engine — nothing installed.
- Some things cannot go in a cloud translator. Unpublished drafts, contracts, medical records, anything under NDA. Every engine here runs on your own machine.
- Bilingual, not replaced. The translation goes under each paragraph, so a sentence that looks wrong can be checked against the original instead of taken on faith.
- PDFs too. Papers and specs open in a bundled viewer where selecting text works the same as on a web page — and a whole paper translates page by page, with the layout left alone.
- And EPUBs. A book opens in a reader of its own, with its contents, its pictures and your place in it kept. Ask for a translation once and it follows you from chapter to chapter.
- Traditional Chinese that reads like Taiwan. Most tools translate into
Simplified and convert character by character, which produces
界面,公裡,數據庫where a Taiwanese reader expects介面,公里,資料庫. OpenRead converts at the phrase level and corrects the vocabulary on the way out.
⬇ Add to Chrome from the Chrome Web Store. One click, and it updates itself.
If you would rather not install from the store,
download the latest release,
unzip it, open chrome://extensions, turn on Developer mode, click
Load unpacked, and pick the unzipped folder. The same package works in
Edge, from edge://extensions.
Either way, that is the whole setup. OpenRead uses Chrome's own on-device translator, whose language pack Chrome downloads once per language pair — about 350 MB, measured at 82 seconds on a 4 MB/s connection. That download starts the moment the extension is installed rather than the first time you press translate, and the extension holds it open until it lands: Chrome does not resume an interrupted one, it waits three minutes and then starts over. After it arrives, a page takes about two seconds.
- Chrome 138 or later, on desktop. The built-in translator is Chrome's, not
this extension's, so a browser without it has nothing to run on. Edge
works — measured on Edge 151, where the translator is there and a cold
first translation takes about twenty seconds — with one difference: Edge has
already claimed the three suggested shortcuts for itself, so it registers
them unbound. Everything is still reachable from the right-click menu and the
popup, and the keys can be assigned by hand at
edge://extensions/shortcuts. Firefox is not supported, and the honest reason is that it has no equivalent API: the only engine left there is Ollama, which means a server and a five-gigabyte model — the opposite of what this is for. - Thirty-nine target languages, every one probed against a real Chrome rather than taken from documentation. The source language is whatever the page is in — that side is not a list.
- Three settings for how it looks: the original beside the translation or the translation alone; a line, nothing, a dashed underline or a tinted background; smaller, the same, or larger. They apply to a page you are already looking at, without re-translating it.
- The first translation of a language pair waits for a download. Chrome fetches the model itself, once per pair per browser profile — 30 seconds to two minutes. The popup says whether yours is ready and offers to fetch it before you need it, so the wait happens while you are not waiting.
- Chrome decides which pairs exist. It downloads packs per language pair, and a pair it has no pack for is one the built-in engine cannot do. When that happens OpenRead falls back to Ollama; with no Ollama configured, that is where the attempt ends.
- Whole page — right-click → Translate this page with OpenRead, or the button in the popup. What you can see is translated first, and the page keeps up on its own as you scroll and as it loads more, so an article or a feed needs one press rather than one per screen. Content inside an iframe — an embedded document, a comment thread — is translated too; an ad slot in one is not, on the strength of what is in the frame rather than where it sits. A badge counts progress and offers Stop. Do it again to put the page back.
- Without asking — in the popup, set Translate automatically to Pages in another language and they translate themselves as they load. Off by default, and any site can be excluded with one checkbox.
- One paragraph — hold Alt and point at it. The paragraph that stopped you, without a drag and without rewriting the page. Remappable to Ctrl or Shift, or off, in the popup.
- What you are writing — press
Ctrl+Shift+Kin any text box and it is replaced with its translation. Its own output language in the popup, English by default.Ctrl+Ztakes it back. - One selection — select text, click the floating 文, or press
Ctrl+Shift+Y. Escape closes the panel. - Names it keeps getting wrong — list them in the popup, one a line.
OpenReadon its own keeps the name as written;bug = 瑕疵pins how a term is translated. The built-in engine takes no prompt, so the term is held out of the sentence and put back after. - A paper — open any
.pdfand it lands in the bundled viewer, where selecting text works exactly as it does on a web page. Press translate and each page's translation is placed under the page itself, so the layout, the figures and the equations stay as the author set them. For a PDF already on your disk, turn on Allow access to file URLs inchrome://extensions→ Details first; Chrome withholdsfile://from every extension until you do. - A book — press Open an EPUB… in the popup, then choose the file or drop it onto the page. Contents down the side, Next and Previous in the toolbar, and your place kept per book, so closing the tab is not losing it. Press Translate once and every chapter after it arrives translated. The book is read out of the file in your hands: nothing is uploaded, and nothing inside it is fetched from anywhere — a picture comes from the archive, and a chapter that points at a remote image simply does not get one.
| The whole page, bilingual | A paper, in the bundled PDF viewer |
|---|---|
 |
 |
Any translation can go into your Obsidian vault as
Markdown, in one tap, through an obsidian://new URI — no extra permissions
and no server:
---
title: 'Ollama'
source: 'Wikipedia'
url: https://en.wikipedia.org/wiki/Ollama
captured: 2026-08-06T03:41:00.000Z
from: en
to: 'Traditional Chinese'
status: raw
tags: [openread]
---
## Original
> Ollama is an open-source software platform…
## Translation
Ollama 是 Jeffrey Morgan 和 Michael Chiang 於 2023 年開發的開源軟體平台…status: raw is a deliberate handoff: a stronger downstream model can pick up
the unprocessed captures, synthesize them, and flip the flag. OpenRead does the
cheap, reliable part on-device and leaves the expensive part to something whose
job it is.
The same pipeline as a command and as an MCP server, so a script or an agent can translate locally too:
npx openread "The quick brown fox jumps over the lazy dog."
cat paper.txt | npx openread --to Japanese
claude mcp add openread -- npx -y openread mcpNot a reimplementation — src/node/translate.ts calls the identical function
the extension calls, and a test pins the CLI's defaults against the
extension's.
| Chrome built-in (default) | Ollama | |
|---|---|---|
| To install | nothing | a server + a 5.2 GB model |
| A 44-block article | ~2 s | 54 s |
| chrF on 27 fixtures | 36.5 | 46.4 |
| Reads surrounding context | no | yes |
| Choose the model | no | yes |
| Capture enrichment | no | yes |
Both run entirely on your machine, and both numbers come from the same 27
fixtures against the same Taiwanese references — the built-in engine's are in
eval/BUILTIN-RESULTS.md, the model's in
eval/BENCHMARK-RESULTS.md. Ten chrF is the price
of installing nothing; a whole segment finishes on the built-in engine in about
the time Ollama takes to paint its first character.
Switch in the popup; ollama pull qwen3 and ollama serve is the whole Ollama
setup, with no OLLAMA_ORIGINS step — here is why.
No account, no API key, no analytics, and no remote endpoint anywhere in the
code. On the default engine the extension makes no network request at all; on
the Ollama engine, only to the server URL you configured, on your own machine.
Permissions are storage, activeTab, contextMenus and
declarativeNetRequest — the last one strips the Origin header from this
extension's own requests to your Ollama server, scoped so a web page's requests
can never match it. Full policy.
That is the author of the extension telling you the extension is safe, which is worth very little. So it is also a test you can run:
pnpm build && pnpm e2e:egressIt loads the built extension into a real Chrome, translates a page served from
127.0.0.1, and records every request every part of the extension makes. It
fails if one left the machine, if one went to Ollama — which would mean the
default engine was not what served the run — or if nothing was translated at
all, since that is the easy way to send nothing and prove nothing.
pnpm install
pnpm build # -> .output/chrome-mv3, load it unpacked
pnpm test # 780 unit tests
pnpm eval # offline reliability eval -> eval/RESULTS.md
pnpm eval:builtin # score the default engine in a real Chrome -> eval/BUILTIN-RESULTS.md
pnpm bench # live model benchmark (needs Ollama)
pnpm shoot # re-shoot every screenshot and the clip on this page
pnpm e2e:page # whole-page translation in a real Chrome, end to end
pnpm e2e:stress # hostile pages: infinite feeds, route swaps, rapid toggling
pnpm e2e:egress # prove the default engine sends nothing, in a real Chrome
pnpm e2e:epub # a real .epub opened, read and translated in a real Chrome
pnpm e2e:first-run # what a brand-new install does on a brand-new profile
pnpm e2e:prefetch # prove installing it starts the language-pack download
pnpm e2e:page-pack # prove opening a foreign page starts its pack too
pnpm e2e:selection # select text, click the 文 icon, read what comes backHow it is built → — the reliability layer, the
streaming assembler, the evals, the model benchmark, and the bugs that produced
each of them. docs/ARCHITECTURE.md has the module map;
CONTRIBUTING.md has the workflow.