crabe

Privacy policy — Text Reliability Assessment (Crabe)

Extension version: 3.0 Last updated: 10 August 2026 Developer: Céline Lenoble — crabe@fastmail.com Source: AGPL-3.0. Everything described below can be checked against the code.

The short version

Crabe has no server. It collects nothing, stores nothing about you, and sends nothing anywhere unless you click Analyse — and then only to services you chose and configured yourself, with your own key and your own billing.

There is no account, no identifier, no telemetry, and no analytics. Nothing is sent to the developer, because there is nowhere for it to be sent.

What leaves your browser, and when

Nothing happens in the background. Every request below is caused by you clicking Analyse this page.

1. The page text, to the AI provider you selected — always

The text of the page you are analysing (up to 50,000 characters), its URL, its title, and the author and date declared in its markup, are sent to the one provider you configured:

Provider	Endpoint	Governed by
Google Gemini	`generativelanguage.googleapis.com`	Google’s privacy policy and Gemini API terms
Anthropic Claude	`api.anthropic.com`	Anthropic’s privacy policy
A local model via Ollama	`localhost:11434`	nobody — it does not leave your machine

Choosing a provider and entering its key is the consent for this. Crabe does not ask twice.

2. A search query, to that same provider — only if you switch it on

Lateral reading asks what independent sources say about the publisher, the author and the page’s central claim. To do that it sends a search query built from those three things through the provider’s own web-search tool.

Off until you turn it on, in Settings. The switch states what is sent and to whom, next to the switch.

3. A request to each link the page points at — only if you switch it on

Citation checking fetches the page’s own outbound links to see whether they resolve. Those sites receive an ordinary request and learn that someone followed the link. Cookies are not sent, so the request is anonymous to them beyond an IP address and the ordinary metadata any visit carries.

Off until you turn it on. Separate from the switch above, because it contacts different people.

A withheld switch is enforced, not discouraged. The request is refused in code before a network socket is opened, and a missing or corrupt setting reads as withheld rather than as granted.

What is kept on your machine

Your settings, in the browser’s extension storage:

which provider and model you chose, and your two consent switches — these use storage.sync, so they follow you to other browsers where you are signed in;
your API keys, in storage.local only. They are deliberately excluded from storage.sync, because that store replicates to Mozilla’s or Google’s servers with browser sync enabled, which would mean handing a credential to a third party you were never asked about. The partition is declared in the code and enforced by a runtime check and an automated test, not left to convention.

Nothing else. The text of an analysed page is never written to disk. It is passed in memory to the panel, used for the analysis, and gone when the panel closes. Results are held in memory only and are forgotten when you close the panel; if you want to keep one, the export button writes a Markdown file where you choose to put it.

Changed in 3.0. Earlier versions wrote the extracted text, URL and title to local storage, where nothing read them and nothing deleted them. That is removed. Version 1.x also stored the Gemini API key in storage.sync; if you used it, opening Settings and saving once removes any key found there.

What Crabe never does

No telemetry, no analytics, no usage statistics, no crash reporting.
No accounts, no sign-in, no identifiers of any kind.
No advertising, no profiling, no selling or sharing of anything with anyone.
No browsing history. Crabe sees one page: the one you clicked Analyse on.
No background or automatic analysis.
No remote resources. Every script, stylesheet, font and image is bundled in the extension. Opening Settings contacts nobody.
No images, audio or video are read or transmitted. Crabe assesses text only.

Permissions, and why each exists

Permission	Why
`activeTab`, `scripting`	To read the text and links of the page you clicked Analyse on, at that moment. Not used anywhere else.
`tabs`	To know which tab a result belongs to, so an analysis is never displayed over a different page.
`storage`	Your settings and keys, as above.
`sidePanel` (Chrome)	To open the panel the analysis runs in.
Access to `<all_urls>`	Optional, never asked for at install, and asked for in two situations. When you click Analyse, for the page in front of you alone. And for all sites, only if you switch on citation checking — checking whether a link works means requesting it, and a page’s links point at other people’s sites, so there is no narrower permission that would reach them. Switching that off gives the broad access back. Leave it off and Crabe only ever holds the pages you have analysed.
`generativelanguage.googleapis.com`, `api.anthropic.com`, `localhost:11434`	The three providers, so a request to the one you chose is possible and a request anywhere else is not.

Your controls

Change or remove your keys at any time in Settings.
Withdraw either consent at any time; the next analysis obeys immediately.
Use a local model and the page’s text never leaves your machine: it is read by a model running on it. Lateral reading is then unavailable, because a local model has no web-search tool — Crabe says so rather than quietly using a cloud provider you did not choose. Citation checking is the one exception, and only if you switch it on: verifying that a link resolves means requesting it, so those requests go to the linked sites whatever model you chose. Leave that switch off and a local model makes the analysis entirely local.
Remove everything by uninstalling the extension; the browser deletes its storage with it.
Data already sent to Google or Anthropic is held under their retention policies, not ours. We never see it.

Children

Crabe is not directed at children and collects nothing from anyone, including them.

Changes

Material changes will be published here with a new date, and the version this policy describes is stated at the top. This document describes version 3.0.

Contact

crabe@fastmail.com

This site is open source. Improve this page.