Stagehand now caches the results of repeated actions, eliminating redundant LLM calls automatically, no code changes required.

When Stagehand encounters a page structure it's seen before, it serves the cached result instantly, no LLM call needed. DOM hashing validates each match, so accuracy stays high even as cache usage scales.

The result: up to 2x faster execution and ~30% cost reduction on repeat workflows. Your agents get more efficient the more they run.

Caching is conservative by default and scoped to your project. It kicks in automatically for all Stagehand users on Browserbase.

Read the docs →

Browserbase is now a one-click integration on the Vercel Agent Marketplace. Your agents get production-grade browsers without running browsers themselves.

Agents deployed on Vercel Sandbox can connect to remote browsers over CDP, keeping your agent logic on Vercel while the browser runs on infrastructure built for automation. One API key, unified billing, no proxies or complex configuration. The integration also works alongside Vercel's AI Gateway and Sandbox out of the box.

We're also launching Web Bot Auth support with Vercel. Browserbase agents now send cryptographically signed requests that Vercel verifies, so your agents can reliably browse Vercel-hosted deployments without getting blocked by security layers.

Read more ->

Functions let you deploy agents and automations directly to Browserbase and run it next to the browser session it controls. You define, deploy, and invoke a function, and Browserbase handles execution, browser lifecycle, and results as one system.

What this means:

• Lower latency, up to 70% faster

• Fewer moving parts and less infrastructure to manage

• Async and long-running executions with results tied directly to real browser sessions.

• Local development mirrors production using real Browserbase sessions.

Available now in TypeScript (documentation)

npx @browserbasehq/sdk-functions init

We’ve completely rebuilt session recordings from the ground up.

Instead of DOM-based replay, Browserbase now records what the browser actually renders using a CDP-based screencast. This makes replays deterministic, accurate, and trustworthy.

What’s new:

• Pixel-accurate screen recordings (no DOM reconstruction)
• Event-driven frame capture (records only when the screen changes)
• Asynchronous, constant-time video encoding using HLS
• Seamless multi-tab playback with a shared global timeline
• Secure, signed segment delivery with CDN enforcement
• Just-in-time encoding to keep costs low and playback fast

If you tried our old session replays and lost trust in them, this rebuild is for you.

Read more here: https://www.browserbase.com/blog/session-recordings

Python, Go, Java, Ruby, Rust...Stagehand is now available on every language! With Stagehand v3 we set course to improve Stagehand's extensibility and to make it easier to bring along the tools that you all love.

Today we're extending this across all languages. Go, Ruby, Java, Rust... and a revamped Python with v3 support.

In order to get here, we are releasing the canonical implementation of Stagehand powered by our newest and most powerful features, from which every SDK will benefit immediately.

This will allow us to move faster, ship more, and focus on what matters most - building the best AI browser automation framework together.

To learn more about this major update:
Check out the docs.
Read the blog post
Use the python Migration guide

Introducing Stagehand v3! 🤘

This marks the beginning of a new foundation for Stagehand, going closer to the metal and providing more flexibility, speed, and control over the browser. We hope this inspires you to help us shape the future of web automation

Featuring:

• A generous 20-40% speed increase across act, extract, & observe calls
• Compatibility with Playwright, Puppeteer, and Patchright
• Automatic action caching (agent, stagehand.act). Go from CUA → deterministic scripts w/o inference
• A suite of non AI primitives:
  • page
  • locator (built in closed mode shadow root traversal, with xpaths & css selectors)
  • frameLocator
  • deepLocator (crosses iframes & shadow roots)
• bun compatibility
• Simplified extract schemas
• CSS selector support (id-based support coming soon)
• Targeted extract and observe across iframes & shadow roots
• More intuitive type names (observeResult is now action, act accepts an instruction string instead of an action string, solidified ModelConfiguration)

With Director, anyone can design repeatable browser agents with just a single prompt in natural language.

Our agent is excellent at fetching you data that isn’t easily scrapeable, such as government registries, multi-page reports, websites that need a login, etc. Using our powerful Stagehand framework, Director can fill nearly any form and automate tedious submission work.

Director works completely hands-free, so you can spawn many of them to operate over thousands of websites, such as municipal websites that vary slightly by each region.

Get started at Director.ai

Project Admins can now change project concurrency without needing to contact us. This can be done from the projects page, project settings, or organization settings!!

Stagehand agent can now call tools and integrate with third-party MCP servers! This increases efficiency for certain tasks (web search, authentication, google workspace) and expands the scope of what our agent can do (db insertions, memory/reasoning, custom tools).

Stagehand TS 2.4.3

Shadow DOM support! Set experimental:true in the Stagehand constructor to try it out

• Fix for same-process iframes
• Enabling scrolling within iframes
• Handle namespaced elements in xpaths
• Bump Zod to be compatible with v4
• Patch for new GPT-5 api format

Stagehand Python 0.5.1

• Shadow DOM support for env="BROWSERBASE" (also requires experimental=True)
• Multi-tab support! S/o Arun
• LLM api_base customization! S/o @Sanveer Singh Osahan
• Fix for stagehand.metrics on Browserbase
• Fix for agent max_steps parameter
• Fix for Litellm dependency errors with OpenAI on newer versions
• Patch automatic download behavior! S/o @Shubhankar
• Patch for new GPT-5 api format