Browserbase Changelog

Develop agents that can browse the web with Browserbase and Inngest.

There are two ways to use Browserbase with AgentKit:

1. Create your own Browserbase tools - If you want to build simple actions on webpages with manual browser control.
2. Use Browserbase’s Stagehand library as tools - If you want autonomous browser and resilient scraping.

Get started today

The v1.14.0 Stagehand release includes performance improvements to the act() method using observe() under the hood, adds targeted extraction via XPath strings in the selector parameter, and introduces support for new AI models like GPT-4.5-preview, Claude 3-7-sonnet-latest, Cerebras LLMs, and Groq LLMs.

See the full release

We’ve added 8 new tutorials to our playground to make it easier to get started and trial features.

• “Enabled AdBlocker” to test out visiting websites with Ads
• “Two Tabs” guide for managing more than one tab
• “GitHub Sign In” tutorial to show how to sign in within Browser live view
• “Training Mode” playground for generating code
• “AI Codegen” playground to test out our AI code gen feature
• “Playwright Console” playground to test our playwright code gen
• “Advanced Tips” section with feature examples

We’re excited to announce our new free plan for side projects & exploration. Users have always been able to test out Browserbase with 10 free browser sessions, but we’re increasing that usage to an hour of session time with 7-day data retention.

Read more (blog) & try it out today (sign up)

Stagehand version 1.13 has been released with the option to customize your local browser.

Configuration highlights:

• Chrome executable path (for pointing it at your local chrome)
• Custom user data directory
• Viewport settings

The release also made updates to Act() which will automatically attempt to self-heal from a stale observe result and shipped a better developer experience around iframes.

See the full release here.

Now supported in Stagehand, Langchain is a Python and Node.js framework to build applications on top of large-language models (OpenAI, Llama, Gemini).

npm run langchain to run our new example (demo).

You can also add Browserbase to your Langchain application (docs).

The Browserbase Playground was built to help you quickly run sessions, generate code with AI, and debug Playwright scripts. The new version of the Playground includes a more intuitive user experience, clarified onboarding language, and an easier user interface to explore the full capabilities of browserbase.

Try it out today at browserbase.com/playground and reach out to hello@browserbase.com if you have any feedback!

Lot of new changes for observe this week. If you set the returnAction flag to true, it’ll return the suggested Playwright method to run.

Also, Stagehand leverages the accessibility (a11y) tree to improve speed, accuracy, and costs fewer tokens. Using the a11y tree, however, means Stagehand may be able to act on items that aren’t directly visible (i.e. hidden behind a dropdown). To use these new features, set onlyVisible: false

const results = await page.observe({
     instruction: "click the sign in button",
     returnAction: true, // Return a Playwright action
     onlyVisible: false // Leverage the a11y tree (faster/better/cheaper)
})

// Act on an observation
// NO ADDITIONAL LLM CALL HERE! Just does the thing :)
await page.act(results[0])

This returns a list like this, which now contains method and arguments

[
  {
    description: 'The button saying sign in...'
    method: 'click', // or another playwright method
    arguments: [],
    selector: 'xpath=/html/body/div[1]/main/section/div[1]'
  }
]

Now, you can cache the selector, method, and arguments if you want to repeat the same Playwright without more LLM calls.

We also made some huge improvements to create-browser-app, making vercel AI SDK + ollama first-class citizens and greatly improving cursorrules.

Normally, sessions begin the shut down process on disconnect. But you can change this behavior so that sessions stay alive till they either timeout or you stop them manually.

Previously, this was reserved for Scale plan customers, but we’ve opened it up for all plans! This makes it possible for any customer to clone our newest open source project, Open Operator.

Read more about keeping sessions alive in our documentation, or view the open operator project & repository

Our List Sessions endpoint has always supported filtering sessions by status, but didn’t give the level of configuration sometimes required. With user metadata, you can now attach arbitrary JSON data to sessions, which then gives you additional flexibility when querying sessions.

Learn more in our documentation: https://docs.browserbase.com/features/sessions#user-metadata

Usage data for current and last billing cycle has now been added to the Browserbase Overview page. You can now see your usage segmented by last 7 days, last 24 hours, or your billing cycle. Data includes total sessions, average session duration, total proxy bandwidth, and average session proxy bandwidth.

When using the Browserbase playground the code will now mute when sessions are running. There is also a new “clear code” button to reset back to fresh state as well as a share button for copying code or sharing the URL.