CloakBrowser patches the wall web agents keep hitting
+1,325 stars in a single day is what GitHub Trending looks like when a real pain point gets solved. CloakHQ released CloakBrowser v0.3.26 in late April; this week the repo crossed 6,012 stars and the reason is not mysterious. Every team building a browser agent — browser-use, Crawl4AI, Stagehand, anyone with Playwright in production — hits the same wall: bot detection. Cloudflare Turnstile, reCAPTCHA v3, FingerprintJS. The agent works locally, breaks in the wild.
What CloakBrowser does that playwright-stealth and undetected-chromedriver do not is patch the actual Chromium source. 49 C++ patches at the binary level, not JavaScript injection or launch-flag tricks. The repo documents 30 out of 30 detection sites passed, including a 0.9 reCAPTCHA v3 score — the score real humans get. You import-swap with Playwright, run the same automation code, ship.
The pricing math is the structural part. If a web agent costs ten cents per task in tokens but fails 40% of tasks at the door because of detection, the LLM is doing free work for a broken pipeline. Solving the detection layer at the source is what changes the unit economics, not bigger models, not better prompts. This is the same insight that drove Web Speed's DOM-to-JSON trick and PageIndex's vectorless retrieval — the LLM is not always the right hammer.
MIT-licensed, drops into existing browser-use / Crawl4AI / Stagehand / LangChain stacks. The velocity says every team running web agents in production is racing to swap it in.
Repo: https://github.com/CloakHQ/CloakBrowser
Site: https://cloakbrowser.dev/
← Back to all articles
What CloakBrowser does that playwright-stealth and undetected-chromedriver do not is patch the actual Chromium source. 49 C++ patches at the binary level, not JavaScript injection or launch-flag tricks. The repo documents 30 out of 30 detection sites passed, including a 0.9 reCAPTCHA v3 score — the score real humans get. You import-swap with Playwright, run the same automation code, ship.
The pricing math is the structural part. If a web agent costs ten cents per task in tokens but fails 40% of tasks at the door because of detection, the LLM is doing free work for a broken pipeline. Solving the detection layer at the source is what changes the unit economics, not bigger models, not better prompts. This is the same insight that drove Web Speed's DOM-to-JSON trick and PageIndex's vectorless retrieval — the LLM is not always the right hammer.
MIT-licensed, drops into existing browser-use / Crawl4AI / Stagehand / LangChain stacks. The velocity says every team running web agents in production is racing to swap it in.
Repo: https://github.com/CloakHQ/CloakBrowser
Site: https://cloakbrowser.dev/
Comments