zendriver-rs

Async-first, undetectable browser automation via the Chrome DevTools Protocol — drive real Chrome from Rust, or hand the keys to an LLM agent over the Model Context Protocol.

A Rust port of zendriver. Drives Chrome via raw CDP — no WebDriver, no JS shim — with anti-detection patches baked in by default.

📖 User guide & full documentation → · 🦀 API reference (docs.rs) → · 🤖 MCP server for AI agents →

Quick example

use zendriver::Browser;

#[tokio::main]
async fn main() -> zendriver::Result<()> {
    let browser = Browser::builder().headless(true).launch().await?;
    let tab = browser.main_tab();

    tab.goto("https://example.com").await?;
    tab.wait_for_load().await?;

    // Find by visible text (auto-waits up to the selector's timeout).
    let link = tab.find().text("More information...").one().await?;
    link.click().await?;
    tab.wait_for_load().await?;

    // Read back from the page's main world.
    let title: String = tab.evaluate_main("document.title").await?;
    println!("title: {title}");

    browser.close().await?;
    Ok(())
}

More working examples in crates/zendriver/examples/.

Feature matrix

Feature	Default?	Use case	Extra deps
stealth	yes	Anti-detection: spoofed UA/platform, isolated worlds, JS shim	(built-in to zendriver)
interception	no	Block/modify requests via CDP Fetch.*; rule-based + streams	zendriver-interception
expect	no	Playwright-style expect_response() / expect_request()	(in-tree, no extra crate)
cloudflare	no	Solve Cloudflare Turnstile challenges	zendriver-cloudflare
imperva	no	Imperva WAF / Incapsula bypass (reese84 / legacy / CAPTCHA)	zendriver-imperva
datadome	no	DataDome bypass (device-check / CAPTCHA / block) + Surface::Webgpu coherence	zendriver-datadome
monitor	no	Passive network monitor — buffered HTTP / WebSocket / SSE events via tab.monitor()	(in-tree, no extra crate)
geo	no	Country code → coherent locale + languages (Accept-Language) persona overlay	(via zendriver-stealth)
tracker-blocking	no	Opt-in third-party tracker / fingerprinter host blocklist (curated bundled list + BYO file / URL)	reqwest + dirs
fetcher	no	Auto-download a pinned Chrome for Testing build	zendriver-fetcher + reqwest/zip

Separate binary crate (not a feature on zendriver):

Crate	Use case	Install
zendriver-mcp	Drive a stealth Chrome from any LLM agent — Model Context Protocol server, 70 tools, stdio + streamable HTTP	cargo install zendriver-mcp (docs)

Install

Pick the use case that matches what you're building.

Just browse:

cargo add zendriver

Default stealth is on.

Stealth scraping (explicit):

cargo add zendriver --features stealth

Same as above — only spell out the feature if you want it visible in Cargo.toml.

Everything:

cargo add zendriver --features "interception expect cloudflare imperva datadome monitor geo tracker-blocking fetcher"

Adds request interception, expect() matchers, Cloudflare Turnstile bypass, Imperva WAF / Incapsula bypass, DataDome bypass, the passive network monitor, country→locale geo overlay, the opt-in tracker/fingerprinter blocklist, and the Chrome for Testing fetcher.

Drive a stealth browser from your AI agent

zendriver-mcp is a first-class Model Context Protocol server that hands the entire zendriver-rs surface to any LLM client — Claude Desktop, Claude Code, Cursor, or your own agent loop. 70 tools, two transports (stdio + streamable HTTP), and the same stealth-by-default fingerprinting baked into the lib. Unlike generic browser MCP servers, this one bypasses Cloudflare Turnstile, ships an isolated-world JS eval that survives anti-bot detection, and lets agents persist auth state across sessions.

cargo install zendriver-mcp

Claude Desktop / Claude Code config:

{
  "mcpServers": {
    "zendriver": {
      "command": "zendriver-mcp"
    }
  }
}

Even easier — the Claude Code plugin:

claude plugin marketplace add TurtIeSocks/zendriver-rs
claude plugin install zendriver@zendriver-rs
# then, in a session:
/zendriver:setup     # prebuilt (no Rust), source, or link

No manual MCP config — the plugin bundles the server plus scraping skills, the /zendriver:scrape and /zendriver:extract commands, and a zendriver-scraper subagent.

What agents get out of the box:

• Stealth navigation — browser_open, browser_goto, browser_back/forward/reload, browser_wait_for_idle, plus a runtime-swappable browser_set_stealth_profile (auto / native / spoof_macos / spoof_linux / spoof_windows)
• Selector-based find + actions — one Selector arg works across browser_find, browser_click, browser_type, browser_press, browser_set_value, browser_upload, etc., with CSS / XPath / visible-text / ARIA-role lookups and per-frame scoping
• Three ways to "see" the page — browser_html (trimmed DOM), browser_screenshot (PNG / JPEG / WebP as inline image content), browser_element_state (visibility / geometry / attrs)
• Stateful primitives agents need for real work — browser_cookies_persist for save/load auth, full browser_storage_*, multi-tab management, frame traversal; browser_monitor_* passive network monitor (HTTP / WS / SSE); browser_request for HTTP from the page's own session
• Anti-bot superpowers (gated cargo features, on by default for the published binary):
  • browser_solve_turnstile — Cloudflare Turnstile bypass without a CAPTCHA-solving service
  • browser_open block_trackers / tracker_blocklist — opt-in third-party tracker & fingerprinter host blocking (curated bundled list + bring-your-own file / URL)
  • browser_solve_datadome — DataDome device-check / CAPTCHA / block bypass (with Surface::Webgpu coherence)
  • browser_intercept_* — block/redirect/respond/modify requests via CDP Fetch.*
  • browser_expect_register / _await — Playwright-style "wait for response/dialog/download" matchers, split across MCP calls so the agent can act in between
  • browser_install_chrome — pull a pinned Chrome-for-Testing build on demand
• Actionable errors — every error carries an _meta.suggested_next hint pointing the agent at the right recovery tool (e.g. ElementNotFound → "try browser_html to inspect")

See the MCP chapter for the full tool reference, CLI flags, HTTP-mode operator notes, and troubleshooting guide.

Phases

Six development phases shipped into the v0.1.0 release. The mdBook covers each surface in depth.

1. Foundation — CDP transport + minimal Browser/Tab/Element. See introduction.
2. Stealth — fingerprint patches + isolated worlds + stealth JS bundle. See stealth.
3. Element API completeness — CSS/XPath/text/role selectors, actionability, input controller, screenshots. See quickstart.
4. Tab/Browser completeness — multi-tab, cookies, storage, frames, nav history, wait_for_idle. See multi-tab + frames.
5. Optional gated features — request interception, expect() matchers, Cloudflare / Imperva / DataDome bypass, network monitor, geo locale overlay, tracker blocklist, Chrome-for-Testing fetcher. See interception, expect, cloudflare, fetcher.
6. Polish + release — trait extraction, rustdoc + mdBook, publish to crates.io.

Comparison

Feature	zendriver-rs	chromiumoxide	fantoccini	headless_chrome	thirtyfour
API ergonomics opinion	builder + auto-wait	raw CDP types	WebDriver verbs	sync wrappers	WebDriver verbs
Stealth out-of-box	yes (default)	no	no	no	no
Multi-tab	yes (first-class)	yes	yes	yes	yes
Interception	yes (Fetch.* wrapper)	yes (raw)	no (proxy-only)	partial	no (proxy-only)
Cloudflare bypass	yes (zendriver-cloudflare)	no	no	no	no
MCP server for AI agents	yes (zendriver-mcp, 70 tools)	no	no	no	no
License	MIT OR Apache-2.0	MIT OR Apache-2.0	Apache-2.0	MIT	MIT
Async runtime	tokio	tokio / async-std	tokio	sync	tokio

Subjective rows marked *opinion. All claims accurate as of the 0.1.0 release; check upstream changelogs before relying on them.

Contributing

See CONTRIBUTING.md. Issues and PRs welcome.

License

Dual-licensed under MIT (LICENSE-MIT) and Apache-2.0 (LICENSE-APACHE) at your option.