/[•]\ Arachne

A fact-checking spider swarm for the browser you already use.

Press Ctrl + Alt + F on any page. Spiders crawl over it, a local AI checks every claim, and the verdicts appear right on the text.

Download · What it does · How it works · Build

---

Why Arachne

• No extension. Arachne reads the page through Windows UI Automation, the same interface screen readers use. It works in Firefox, Brave, Chrome and Edge, with your tabs, your logins, your pages. Nothing to install in the browser.
• The brain stays on your PC. Claims are judged by your own model in Ollama. No account, no API key, no cloud.
• It shows its work. Every verdict comes with a reason, the evidence it used, and a link you can open. Every point of the trust score has a reason too.
• It reads the whole page. The spiders do a quick pass first, then keep reading where you are looking, until the page is done.

What it does

Checks claims	Picks the statements worth checking (numbers, dates, records, quotes, health and science claims) and gives each one a verdict: true, mostly true, mixed, misleading, false, can't confirm, or opinion.
Finds evidence	Searches Wikipedia, and opens the sources the page itself cites to see if they really say what the page claims.
Double-checks "false"	Before anything is called false, a second, skeptical pass looks for reasons the first answer could be wrong.
Briefs you	Summary, key points, what the page doesn't tell you, red flags (clickbait, hidden ads, one-sided framing), and tips.
Trust score	0–100, built from rules you can read: false claims, red flags, missing author or date, selling something.
Reads on	After the quick check it reads the rest of a long page, piece by piece. What's on your screen goes first. Read paragraphs get a silk thread in the margin, and the panel shows a page map with read %.
Ask the page	Ask a question. Arachne searches the whole page for the right paragraphs, answers, and points to them. Click a ¶ to jump there.
Check a selection	Select any sentence and press Ctrl + Alt + C.
The spiders	A small swarm with names like rusty crouton. Their legs land on real words, they grab links on amber threads, say what they found in thought clouds, and glitch the page the way the video that inspired this does.

See it

It catches myths. Seven myths, seven false labels, and a trust score of 3:

Wikipedia, while the spiders read on. Verdicts on the text, silk in the margin, the page map in the panel:

A real capture of the swarm:

Getting started

1. Install Ollama and pull a model, for example ollama pull qwen3:14b. Arachne uses the first installed model unless you pick one in Settings.
2. Download Arachne, unzip it anywhere, and run Arachne.exe. A spider appears in the tray.
3. Open a page in your browser, click into it, and press Ctrl + Alt + F.

Keys	What happens
Ctrl + Alt + F	Check this page. Press again to put the spiders away.
Ctrl + Alt + C	Check the selected text.

Speed. On an RTX 4080 with a 27B model, a full check of a long Wikipedia article takes about 45 seconds and reading the whole page about 4 minutes. A model only runs fast when it fits in GPU memory: close games before a check. Arachne warns you when the GPU is full.

Settings (gear in the panel, or the tray menu): model, memory size, claims per page, number of spiders, Wikipedia lookups, reading cited sources, double-checking, deep thinking, auto-scroll, keep reading, wild effects, start with Windows.

How it works

flowchart LR
    B["Your browser<br/>(Firefox, Brave, Chrome, Edge)"] -- "UI Automation:<br/>text, layout, links" --> R["Reader<br/>paragraphs from layout,<br/>exact scroll tracking"]
    R --> BR["Brain<br/>claims, brief, verdicts"]
    BR <--> O["Ollama<br/>(local model)"]
    BR <--> E["Evidence<br/>Wikipedia + the page's<br/>own cited sources"]
    BR --> P["Panel<br/>(Qt)"]
    BR --> S["Spider overlay<br/>(Direct2D + DirectComposition,<br/>click-through)"]
    S -- "what's under each foot?" --> R

Loading

• Reader. One UI Automation round trip copies the page's text pieces with their positions. Paragraphs are rebuilt from the layout, because Firefox and Chromium disagree about structure. A paragraph's on-screen position, which the browser reports even off screen, gives the exact scroll.
• Brain. The page goes to the model once. Claims and the brief share Ollama's cached prompt. Each claim gets its evidence, a verdict with a reason, and a skeptical second look if it says false. Answers are forced into JSON schemas, so the panel never has to guess.
• Spiders. Eight legs each, with two-bone inverse kinematics and a tetrapod gait. Each foot asks the browser which word is under it and grabs a corner of that word. They are drawn on the GPU in a transparent, click-through window that follows the browser.

Privacy

• The page text goes only to your local Ollama.
• If Wikipedia lookups are on, short search words go to wikipedia.org.
• If "read cited sources" is on, Arachne opens the links the page itself cites next to a claim.
• Both can be turned off for a fully offline check. Nothing else leaves your PC.
• Settings, the result cache and a log live in %LOCALAPPDATA%\Arachne.

Good to know

• The AI can be wrong. Arachne shows its evidence so you can check it: click the sources.
• Can't confirm means no evidence was found. It does not mean false.
• Pages that hide their text from screen readers (some canvas apps, PDFs in the viewer) can't be read.
• Windows only, for now.

Build from source

Needs Visual Studio 2022 (C++ workload), CMake and Qt 6.10 for MSVC 2022 64-bit.

.\build.ps1          # build\ for compiler output, dist\ for the program
.\build.ps1 -Clean   # start fresh

Edit the Qt and Visual Studio paths at the top of build.ps1 if yours differ.

Test hook: Arachne.exe --check "<part of a browser window title>" --claims 4 --ask "question".

Credits

• The look comes from Web crawlers by @rybinfx (Slava Rybin), a beautiful JS + CSS piece. Arachne re-creates the idea in C++ and gives the spiders a job.
• Built with Qt (LGPL-3.0, see THIRD_PARTY.md), Ollama and the Wikipedia API.

License

MIT