Last updated: 22 September 2026

Context Reader has no server. There is no account, no sign-in, and no backend belonging to
this project. Everything below can be checked against the source in this repository.


Nothing. No analytics, no telemetry, no usage statistics, no error or crash reporting.
No identifier is generated for you, and nothing is sent to the developer — there is nowhere
for it to be sent to.



The word you selected and the sentence around it are sent to the Google Gemini API using
your own API key. This is the entire function of the extension.

Google is the data controller for that request. Its handling is governed by the terms you
accepted when you created the key:

• Gemini API Additional Terms of Service
• Google Privacy Policy

Note that the free tier of the Gemini API and the paid tier differ in whether your prompts
may be used to improve Google's models. Which applies to you depends on your own key, not
on this extension. If that matters to you, read the terms above before choosing a plan.


Every typeface the bubble uses ships inside the extension as of 17 September 2026:
Inter for the interface, in Latin, Cyrillic and Greek, and a Noto face for each of the 21
other scripts. Nothing is fetched for them, ever, on any site.

The exception is Chinese, Japanese and Korean. Those families are megabytes each, so they
are still requested from fonts.googleapis.com and fonts.gstatic.com — only when you
open a bubble on a page, and with no referrer, so Google's font servers see your IP
address and that a font was requested, but not which page or site you were on. No word
you look up and no page content is included. Every other language fetches nothing.

Before 15 September 2026 this was worse: fonts were requested on every page load, with the
page's origin attached. That was fixed as soon as it was measured, and bundling removed the
rest.


manifest.json grants exactly one host permission:
https://generativelanguage.googleapis.com/*. Chrome enforces that list. The extension
cannot contact any other server for data, and there is no code in it that tries.


All of this stays in your browser's local extension storage. None of it is transmitted.

| Stored | What it is |
|---|---|
| geminiApiKey or geminiApiKeyEnc | Your API key — plain, or encrypted if you set a passphrase |
| targetLanguage | The language you chose |
| modelPref, modelListCache, modelCooldowns | Which Gemini model to use, and which are rate-limited |
| promptOverrides, promptHistory | Prompts you edited, and their version history |
| savedWords | Words you explicitly saved |
| localLookupCount | A counter shown in the popup |
| latencyLog, latencyDaily | Response times per model, for the speed table in Settings |

The latency records deliberately do not include the words you looked up. They hold a
model name, a timestamp and a duration.

Uninstalling the extension deletes all of it. You can also clear the key at any time with
Revoke in Settings.

• Stored with chrome.storage.local, never chrome.storage.sync — sync would replicate
  it to Google's servers.
• Read only inside the extension's service worker. It is never handed to a content script,
  so no web page can reach it.
• Shown masked in Settings (AIza…3f2) after saving, never in full.
• Never written to logs.
• Optionally encrypted at rest with AES-GCM, using a key derived from a passphrase via
  PBKDF2. The passphrase is never stored, so there is no recovery — forgetting it means
  entering the API key again.

Being straight about the limit of this: hardening protects against page scripts and
accidental exposure. Only the passphrase option protects against someone who can read your
browser profile off the disk. An extension cannot hide a key from software already running
as you, and we will not claim otherwise.


The extension and the website are different things, and only one of them counts anything.

The extension collects nothing. No analytics, no telemetry, no account, no identifiers,
no crash reports. That does not change.

The website counts visits and a few clicks, with two services. Cloudflare Web Analytics
counts visits. Umami counts visits and a few clicks: the install and "Get a Gemini key"
buttons (with which browser the page detected), "Copy for LLM" and "View as Markdown", and
the language chosen in the home page's demo. Umami's script comes from cloud.umami.is and reports to gateway.umami.is.

Neither service sets a cookie or stores anything in your browser, and neither follows you
between sites. Between them they record the page, the referrer, the browser, operating system
and device type, and a coarse location (the country) — so we can tell whether anyone is
reading and what leads to an install.

They are the only things the site loads from anywhere else: no other trackers, no advertising,
no embedded video, and the typeface is served from the site itself.

For one day, 21 to 22 September 2026, the site also offered Microsoft Clarity, a tool that
records how a page is used, and loaded it only for visitors who clicked Allow. It has been
removed. If you allowed it, Clarity's cookies on the site are no longer read by anything and
expire on their own; clearing the site's data removes them now.

The site itself remembers one thing: the language you pick in the home page's demo, kept
in your browser's local storage so it is still chosen next time. It never leaves your device,
and clearing the site's data in your browser removes it.

Nothing about the extension, and nothing you look up, ever reaches the website: a lookup
happens in your browser and never touches this site.

This page is also published at
contextreader.github.io/privacy. Where the two
differ, this file — which is version-controlled and public — is the one that counts.


Not directed at children under 13. The extension collects nothing from anyone; the website counts visits as described above.


Material changes will be noted in the repository history and the date above updated. The
history of this file is public.


Email contextreader@gmail.com, or open an issue on the
project's GitHub repository.

Questions about how Google handles the lookups themselves go to Google — this project
never receives them.

https://contextreader.github.io/privacy