An agent browser that actually allows your agent to browse the web without getting bombed with bot detection and Its also focused on token efficiency, (Rust, OSS)

So, hear me out, Every agent browser I've seen came with 15-20 tools. Playwright MCP alone is 13,700 tokens of definitions. Your agent burns more tokens reading the tool docs than actually browsing. So I tried the opposite. 5 tools. One Rust binary. No Node.js. npm install -g bladebro && bladebro mcp first command in CLI auto-starts a daemon, Chrome stays alive across all calls: bladebro nav whatever.com bladebro see content bladebro act click e5 bladebro stop Same engine behind both. ~1,900 tokens of tool definitions total HAS MCP and CLI support The 5 tools act — click, type, fill, scroll, press, hover, navigate, wait, eval, download, upload. Plus batch: fill 5 fields and submit in one call instead of 11 round trips. Every action returns what changed, not the full page. A click is ~60 tokens vs 2,000+ with Playwright. see — read the page. Interactive elements with refs, clean markdown, heading outline, or auto-extract: pulls structured data from list pages with zero CSS selectors. Works on Amazon (price/ratings), Reddit (scores), GitHub (stars/forks), HN, Wikipedia. state — cookies, localStorage, tabs, session save/load. Log in once, save the session, restore it next time. run — batch with branching and loops. Multi-step workflows in one call. vision — screenshot with optional numbered ref overlays. Last resort — the structural model is cheaper and more reliable. Re-render immunity Every element gets a structural fingerprint, a hash of its ancestor chain, tag, identity attributes. React destroys and recreates the node? Fingerprint matches, reference survives. Your agent sees "re-render survived" and keeps going instead of dying on "element not found." I checked every major agent browser I could find. All of them lose refs on re-render, Nearly all of them. This is the biggest reliability gap in the space. Bot detection 6 layers, all on by default. Bezier mouse paths with overshoot. Real movementX/movementY deltas on every event — missing these is an instant PerimeterX flag and most tools don't bother. Micro-tremors before clicks. Log-normal typing cadence. Idle drift. No Runtime.enable (that's how DataDome catches most bots). CDP over pipe, zero listening ports. Detects your actual GPU via lspci and spoofs WebGL to match. An AMD machine shows up as AMD, not Intel. Tested live on Zillow and Fiverr, both PerimeterX/HUMAN protected. Full access, no blocks. Sannysoft passes clean. CreepJS shows 0% headless. What it can't do Cloudflare Turnstile needs actual challenge solving, not fingerprint work. You get a blocked: verdict, not a hang Captchas, deliberately — you get a verdict, hand off to a solver Browser extensions — CDP doesn't support it macOS, Windows, and ARM64 Linux binaries are cross-compiled from Linux Apache-2.0, open source. GitHub: github.com/dondai44423/bladebro npm: npm install -g bladebro If you run into issues, comment here or open a GitHub issue. I've tested what I can but edge cases always hide.

添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论