Most AI agents that need the live web break on three fronts: sites block headless browsers, JS-level stealth shims leak tells, and raw HTML is token-heavy for LLM-driven workflows. This project addresses all three by packaging a Camoufox browser engine (Firefox fork that spoofs fingerprints at the C++ layer) behind a small, agent-oriented REST API that returns compact accessibility snapshots and stable element refs instead of raw HTML.
What Sets It Apart
- C++-level anti-detection: spoofs hardwareConcurrency, WebGL, AudioContext, screen geometry and WebRTC before any JavaScript runs — so there are no shim-based artifacts for detection scripts to fingerprint. This reduces many false positives that stealth plugins can't avoid.
- Agent-first primitives: accessibility snapshots are ~90% smaller than full HTML for token efficiency; stable element refs (e1/e2/...) let agents interact deterministically without brittle selectors. So LLM-driven agents spend fewer tokens and fewer steps to complete tasks.
- Operational features for real agents: session isolation and cookie import let agents reuse authenticated contexts; proxy/backconnect + automatic GeoIP keeps fingerprint fields consistent with the exit IP; structured JSON logs and low idle memory targets make it easier to run alongside other services.
- Practical integrations: built-in YouTube transcript extraction (yt-dlp fallback), Docker/Fly/Railway deploy patterns, and tooling to limit resource use make it ready for agent deployments rather than just ad-hoc scraping experiments.
Who It's For and Tradeoffs
Great fit if you run autonomous agents that must interact with real websites reliably (search, login flows, form interactions, or extracting dynamic data) and you need fewer tokenized bytes and more deterministic actions. It also suits small- to medium-scale hosted agent stacks where sharing a single box for multiple agents matters. Look elsewhere if you need extreme-scale web crawling across millions of pages (the project focuses on agent interactions, not bulk scraping throughput), if you require a browser with an explicit vendor support SLA, or if you have strict compliance concerns that preclude using a forked browser binary. Expect a download of native Camoufox binaries (~hundreds of MB) and an ongoing cat-and-mouse with advanced detection vendors — nothing fully eliminates risk of future fingerprinting techniques.
Where It Fits
Compared to Playwright/Puppeteer plus stealth plugins, this approach moves the fingerprint surface into the browser binary itself (so JS-based detection has less to observe). Compared to simple headless Chrome, it trades wider community tooling for stronger anti-detection at the cost of managing a custom browser binary and the legal/ethical considerations of bypassing site protections.
