A browser that lies in C++, and agent skills to prove it

What it does

This repo isn’t the browser — it’s the curriculum. Four Agent Skills (installable as a Claude Code plugin or individually via the skills CLI) teach an agent to drive AntiBrow: writing scripts against the JavaScript or Python SDK, launching the browser directly over MCP tool calls, fanning one scraping task across many logged-in profiles, and auditing whether all of it is actually isolated. Each profile is a persistent identity with its own frozen persona, cookies, and proxy exit IP.

The interesting bit

Most stealth tooling monkey-patches JavaScript from the outside, and anti-bot vendors fingerprint the patches themselves — the patch is the tell. AntiBrow answers canvas, WebGL, audio, fonts, and timezone from inside C++/Blink, so there’s no injected script to find and Web Workers return exactly what the main thread does. Better still, the multi-account-isolation skill teaches an agent to verify the disguise — timezone vs exit IP, canvas stability across relaunches, one GPU across three interfaces — rather than assume it, and it states plainly which layers browser isolation cannot cover at all.

Key highlights

• Personas are sampled from one real machine (30+ categories, 500+ parameters) and frozen at profile creation — no AMD renderer sitting next to an Intel vendor string.
• Timezone, locale, and WebRTC identity follow the proxy: the exit IP is resolved through the proxy before launch, and proxy auth happens in the network stack, not in an extension.
• One profile format shared by the Node and Python SDKs (~/.anti-detect-browser/), so an identity created in one is launchable from the other — and from the desktop app.
• Scraping recipes return JSON per site (Amazon, Reddit, X, GitHub, PyPI, and more) with no selectors to maintain; third-party recipes are fenced in by a network-layer domain allowlist plus SHA-256 pinning.
• Security-minded defaults throughout: secrets only via environment references, pinned versions, and page content treated as untrusted input to guard against indirect prompt injection.

Caveats

• The kernel is closed-source and gated by an API key with plan-based, kernel-enforced concurrency caps; the MIT license covers only the SDKs. This repo is essentially well-curated glue around a commercial product.
• No Alpine/musl builds yet.
• The README is candid that some sites still demand a clean residential exit — consent walls and bot challenges aren’t fixable in parsing code.

Verdict

Worth a look if you run multi-account operations, region-locked ad or pricing checks, or want an agent that operates — and audits — its own browser. Skip it if you need a fully open-source stack or a lightweight headless scraper; the value here rides on the proprietary kernel.