A browser extension that tells a reader how well the article in front of them holds up: which facts are actually sourced, which are checkable and merely unsourced, which the evidence undermines, and where the wording is doing work the evidence does not support.
It reads the page you are on, audits it against a structured grid, checks the audit's own factual objections against the web, and returns a score out of 100 with the weaknesses it found and the exact sentence each one refers to. Findings are highlighted in the article itself, and hovering one in the panel connects it to the passage it came from.
Chrome and Firefox, Manifest V3. Bring your own API key: no backend, no account, no telemetry. Readers who would rather not hold an API key can sign in instead: the hosted mode runs in production on squiggle.fr.
Without the extension, on a phone for instance, the website runs the same analysis on
an article's text pasted into its analysis page (/analyser), and draws the same
report over it. The site installs as an app on Android and iOS, and on Android an
article can be shared straight to it. It runs in hosted mode, on the same account,
free analyses and subscription as the extension.
For teachers, the classroom mode (/classe) prepares an analysed article for class: the
teacher sees every answer alone, the projected board reveals the findings one by one,
and a student sheet and an answer sheet print on A4. Students use no AI and need no
account.
Someone reading a news article, not the journalist who wrote it. A reader cannot edit the piece, so the extension never offers rewrite advice or a pre-publication verdict. It explains how solid the piece is and why, and leaves the conclusion to the reader.
The starting point was a set of industrialised article-review prompts by @sjowall69, which he generously shared with me when I asked. The idea of holding an article to explicit, weighted criteria rather than a vague impression comes from there, and I am grateful for it.
What this extension does has since diverged, deliberately. It addresses a reader instead of an author, so the criteria that graded editorial craft were dropped; framing and rhetorical technique became a domain of its own; assertions are checked against evidence rather than judged from memory; and the score is derived from the findings instead of being volunteered alongside them. None of that should be read as representing his work, and his document is private and not reproduced here.
Five weighted domains, summing to 100, chosen for a reader judging trustworthiness rather than an editor grading craft:
| Domain | Weight |
|---|---|
| Robustesse factuelle et sourcing | 35 |
| Solidité logique et argumentative | 25 |
| Cadrage et procédés rhétoriques | 25 |
| Déontologie et transparence | 10 |
| Soin de la langue | 5 |
Framing and rhetorical technique carries as much weight as logic, because a piece can be factually defensible sentence by sentence and still mislead: a headline overstating what the body supports, loaded wording standing in for evidence, opinion carrying the grammar of established fact, selective quotation, a real disagreement presented as settled, or authority invoked without ever being named.
The composite maps to one of four bands - solide at 80 and above, perfectible at 70, fragile at 60, problématique below that.
Findings are classified as a logical fallacy, an unsupported claim, an overreach, a missing source, a framing problem, or a strength worth keeping.
Each factual objection the audit raises is checked, and the result describes what the article said - never the audit's rebuttal of it. Four states, from the method:
| State | Meaning |
|---|---|
| Vérifiée | evidence was consulted and it backs the article |
| Non sourcée dans l'article | checkable, the article simply does not source it |
| Douteuse | evidence casts doubt on the article's statement |
| Non vérifiable telle qu'écrite | cannot be checked as worded |
The middle two are the distinction that matters. A claim nobody sourced is a sourcing observation; a claim the evidence undermines is a defect. Treating them alike tells a reader an ordinary claim is shaky.
When evidence confirms the article, the audit's objection was wrong. It is withdrawn rather than published as a fault, and recorded with the sources that cleared it, so a real disagreement is visible instead of silently deleted.
A model asked for a mark and a list of defects will happily supply both and let them disagree - describing a fabricated poll accurately in its own weakness note and still scoring that domain generously. So the mark is not taken at face value: every defect the audit raises removes a share of the marks it awarded in that domain, scaled by severity and by how well the defect stands up.
That is a rate, not a ceiling. Defects compound, a badly sourced piece approaches zero on its own, and two articles judged differently stay apart afterwards - the distinction a cap would flatten. Evidence only ever pulls a mark down.
In this mode there is no backend and no intermediary. You supply a key for one of four providers and calls go from your browser straight to them:
- Anthropic
- OpenAI
- OpenRouter
- Google Gemini
The key is encrypted with AES-GCM and stored in your browser's extension storage. It leaves your machine only as a request to the provider you chose. With no key configured the extension says so and stops - it will never substitute a sample report for an analysis of the page you are looking at.
See SECURITY.md for the data flow and how to report a vulnerability.
Holding an API key is a real barrier: you need an account with a model provider, a payment method on it, and enough of an idea of what a token costs to be comfortable. Most readers of a news article have none of that, and asking them to get it is asking them not to use the tool.
So there is a second mode, in production on squiggle.fr: sign in, and the analysis runs on our infrastructure against credits, three free, then a small monthly subscription. It is a choice, not a migration. BYOK stays exactly as it is, remains the default, and is what the extension does with no account at all.
What changes, stated plainly, because it is the whole trade:
- the article text goes to our server, and from there to the configured model provider (OpenRouter by default; Gemini is a deployment setting away). It is used for the audit call and dropped when that request ends; no table in our database holds it.
- the finished report is cached and shared, so a second reader of the same article gets it for free and never sends the article anywhere.
- our server never fetches the article itself — your browser supplies it, from the page you are already reading. A paywalled article is analysed as far as you can read it, and flagged as partial. A YouTube video is analysed from its captions, which the extension reads in your tab; the server never fetches a video or its captions.
- on the website, the text is what you paste, and it travels exactly like the article the extension reads. An address given beside it is kept as a label and a cache key, never requested.
PRIVACY.md has the retention table and the subprocessors, SECURITY.md the controls and the residual risks, and ARCHITECTURE.md the topology.
npm install
npm run build:chrome # -> dist/chrome
npm run build:firefox # -> dist/firefoxChrome: open chrome://extensions, enable Developer mode, choose Load
unpacked, and select dist/chrome.
Firefox: open about:debugging#/runtime/this-firefox, choose Load Temporary
Add-on, and select any file inside dist/firefox.
Then open an article, open the side panel, add your key via the gear icon, and run the analysis.
src/ the extension
crypto/ AES-GCM key encryption and local storage
content/ article extraction, range tracking, in-page highlight overlay
background/ service worker, tab state, analysis lifecycle
sidepanel/ the panel application
ui/ design system, gauges, finding cards, evidence display, onboarding
adapters/ translation between engine, content-script and UI vocabularies
messaging/ typed cross-context messaging and the shared RPC contract
shared/src/ imported by BOTH the extension and the server
engine/ criteria, prompts, research, reconciliation, scoring
client/ provider clients (OpenAI, OpenRouter, Gemini)
auth/signing.ts the request-signing protocol for hosted mode
server/src/ the hosted backend (Cloudflare Workers + Neon)
auth/ magic link, Google OAuth, extension tokens, signature checks
analyze/ the five analysis stage endpoints and the shared cache
billing/ Stripe checkout, portal, webhook, quotas
index.ts the config-driven LLM provider gateway (LLM_PROVIDER)
web/ the public web app (Vite + React), served by the same Worker
src/pages/ sign-in, account, legal pages, and Analyser: paste an article
src/paste.ts a paste as the engine's blocks, and the share target's input
sw.js the installed app's service worker: the app shell, nothing else
shared/ is not a convenience. The moment the same analysis can run in two places, a
duplicated prompt or scoring rule is a latent divergence between what a BYOK reader
and a hosted reader are told about the same article.
ARCHITECTURE.md covers the pipeline and why it is ordered the way it is. docs/product-spec.md is the original product spec.
In BYOK mode nothing reaches a server of mine, because there is no server of mine: the article and your key go straight from your browser to the provider you chose. Hosted mode is the deliberate exception, and it is opt-in and account-gated. PRIVACY.md accounts for every byte that leaves in either mode, and SECURITY.md covers how the key is held.
CONTRIBUTING.md has the build, the gates and the conventions. Releases are documented in docs/RELEASING.md. Deploying the hosted backend and the web app is documented in docs/DEPLOYING.md.
Copyright (C) 2026 TheGlitching.
Squiggle is free software, licensed under the GNU Affero General Public License,
version 3 only (AGPL-3.0-only); the full text is in LICENSE. In short:
you may use, study, modify and redistribute it, and anyone who runs a modified version
as a network service must offer its users the corresponding source code.
Squiggle was first published under the MIT licence. As its sole author, I now offer
the whole codebase, earlier versions included, under AGPL-3.0-only. A licence once
granted cannot be withdrawn, so a copy of an earlier version that someone already
obtained under MIT keeps the MIT terms for that version. Everything published from
this change onward is available under AGPL-3.0-only only.