Repository navigation
guides ai agents
The newest category here, and the one with the most room to grow. An AI agent that drives a browser inherits every fingerprinting surface a human-scripted session does, plus one of its own: the rhythm of an agent's actions is not a human's rhythm, and that is checkable independently of any fingerprint work. Each page here is verified against the framework's actual source before anything is claimed about it - several candidates that looked like a fit by star count turned out not to use Playwright, or not to touch Firefox, at all.
These pages are for developers wiring an agent to this engine. The agent-experience side - what to run, which agent products compare how, what to check when an agent gets blocked - lives on the invisible_playwright_mcp wiki, the agent built on this same engine.
- Give a LangChain agent an invisible_playwright browser - hand PlayWrightBrowserToolkit a launched browser so its tools inherit a real fingerprint.
- smolagents: hand the agent an invisible_playwright tool - smolagents' vision browser drives Chromium via helium, so register a custom tool instead.
- Stagehand and stealth: why a Firefox engine won't drop in - Stagehand is TypeScript and Chromium-channel only, so a Python-launched Firefox cannot drop in.
- crawl4ai stealth mode and custom browser engines - crawl4ai takes browser_type firefox but no executable_path, so you get Playwright's managed build.
- Give an MCP browser server a stealth Firefox engine - point Playwright MCP at a patched Firefox with browserName plus executablePath, minus the profile.
- Back a computer-use agent with a real browser engine - implement the screenshot-and-click Computer interface over an invisible_playwright page, and its honest limits.
- Computer-use agents and browser fingerprint detection - clicking by pixel makes driver flags moot; the engine fingerprint and action rhythm stay checkable. (Moved to the invisible_playwright_mcp wiki; the page here points across.)
- DOM-reading vs screenshot agents: which stealth helps - which stealth signal helps depends on whether the agent reads DOM or clicks pixels.
- AI browser agents and stealth: what fits and what does not - most agents drive Chromium over CDP, so a stealth Firefox does not drop in.
- browser-use gets detected: what you can and cannot change - browser-use drives Chrome over CDP: what BrowserProfile changes, what stays out of reach. (Moved to the invisible_playwright_mcp wiki; the page here points across.)
- Why AI browser agents have their own timing signal - a think-act loop emits machine-regular gaps and instant pointer jumps no fingerprint fixes. (Moved to the invisible_playwright_mcp wiki; the page here points across.)
- AI agent retry loops trip rate limits, not fingerprints - retry and re-plan loops multiply requests into a volume signal; throttle in the agent loop. (Moved to the invisible_playwright_mcp wiki; the page here points across.)
-
Playwright 1.62 Ships MCP In the Box - the MCP server and a new playwright-cli are now bundled into core, runnable via
npx playwright mcpwith no separate install.
- Running an AI browser agent headless on a server - the engine covers the GPU-less headless fingerprint; IP reputation stays yours to supply.
- Run parallel browser agents with distinct fingerprints - each agent gets its own reproducible fingerprint, but one shared exit IP still links them.
- Give each AI agent a reproducible browser identity - one seed keeps a repeated task on one stable device fingerprint.
- Give a browser agent a persistent logged-in session - reuse saved storage_state or profile plus a pinned seed to stay logged in.
- Feed invisible_playwright pages into a RAG index - fetch JS-rendered gated pages as real HTML, then chunk and embed the text.