Changelog

Changelog #001 | 2025-02-09Copied!

Happy Super Bowl Sunday 🏈 Before we settle in to lose money on our respective betting apps, we have some updates we NEED to tell you about.

⭐ New

Introducing Surf.new

Surf.new is an open-source playground for chatting with different web-agents via Steel. We want it to serve as a resource for the community to try out new web agents as they become available, helping developers evaluate what works best for their use-cases. Currently, you can try browser-use's web agent and a browser-specific fork of Claude’s Computer-use agent. We'll be actively maintaining it while using it as a launching pad to showcase new Steel features and improvements in web automation.

It’s pretty neat if you ask us 🤠 Give it a try and let us know what you think! Contributions are more than welcome too :)

Finding Flights on Surf

Embedding and interactive with Live Sessions

Our debugURL has some new life blown into it and, oh boy, are we ever pumped about it. The debugURL you get returned when creating a session is most commonly used for viewing what’s happening live in the browser. A common use-case is embedding that view into your app, such that people can see what’s going on in the browser as actions are being taken. It’s what powers our live session viewer in the Steel dashboard.

In addition to a complete refactoring to improve performance, some of the improvements include:

  • Ability to let a viewer interact with the browser sessions directly via clicking, typing, scrolling, etc. This was a big one lots of people have asked for to powering human-in-the-loop features (think “take control” in OpenAI’s Operator).

  • Ability to show/hide mouse on screen

  • Show/hide URL bar & to toggle be light mode or dark mode

All of which can be turned on or off via UTM params. Check out the docs for more on this here!

Available on Steel Cloud and available soon on the steel-browser repo.

Dimensions for sessions

We now support the ability to set screen + viewport dimensions when creating a session (POST/v1/sessions).

from steel import Steel

client = Steel()

session = client.sessions.create(
    dimensions={
        "width":1280,
        "height":800
    }
)

This helps save you from having to set page viewport on every page load. Which can cause buggy resizing behaviour with your sessions.

Available on Steel Cloud and the steel-browser repo.

Ad blocking

You can now block ads from rendering in your sessions. This is useful for saving on proxy bandwidth, simplifying action space for agents (so they don’t have the option of clicking on ads), and generally speeding up page load times.

It’s defaulted to true when starting a session but you can explicitly turn it on/off by passing a bool into the blockAds param in the create session endpoint (POST/v1/sessions) or via the SDK like so:

from steel import Steel

client = Steel()

session = client.sessions.create(
    block_ads=true
)

🔧 Fixes/Improvements

Lots of bug fixes and improvements across the board including:

  • Fixed with inability to view sessions where proxy was enabled

  • Better scrape errors

  • Improved Proxy usage tracking

  • Fixed multiple issues with Recording sessions

  • Implemented graceful shutdowns

  • Carious dockerfile optimizations

  • Custom executable paths for local browsers when running steel-browser repo locally

💖 First time contributors

Special thanks to the following new contributors to steel-browser who've made the above improvements possible 💖
@marclave, @krichprollsch, @BrentBrightling , @Envek, @danew, @raymelon, @21e8, @QAComet, @mislavjc, and @Emmanuel-Melon

As always, thanks for testing out the Steel beta. We really look forward to more of your feedback and continuing to build with this awesome, curious, and supportive community.


Changelog #000 | 2024-12-20Copied!

Happy AGI day (?) & inaugural changelog post. I wanted to share some updates we made to Steel over the last few weeks!

⭐ New:

  • Just launched a new MCP server for Claude Desktop integration that lets Claude visually browse & interact with the web using our Browser API and Web Voyager

  • Works with self-hosted, local, and cloud implementations

  • Built on a custom Web Voyager implementation using Bounding Boxes and Custom tools to map webpages into LLM action spaces

  • Check it out: https://github.com/steel-dev/steel-mcp-server

🔧 40+ Bugfixes like:

  • Fixed billing page + usage tracking issues

  • Resolved rate limit errors

  • Resolved session creation bugs with large request volumes

  • Fixed compatibility issues with Windows for OS repo

  • Improved recording extension handling, so it should be more stable now

  • Resolved session viewer crashes for sites >10MB

  • Fixed performance issues with loading certain sites in the open source repo

Improvements:

  • Large, complex sites now render in just a few seconds (instead of forever like before)

  • Session launch time improved by ~30%

  • Faster and clearer error messaging

  • Enhanced session viewer stability

🏡 Housekeeping:

  • We're rolling out a Research Grants program to support AI researchers with Steel Credits! If you're working on web scraping research or exploring new ways for agents to use the web, reach out to research@steel.dev and we'd love to support you.

  • We're hiring across engineering roles! Looking for full stack, Applied AI, and infra developers who love open source, AI, and tackling challenging problems. Plus, there's a $5,000 referral bonus if you help us find the right person! More details: https://steel-dev.notion.site/jobs-at-steel

Thanks for testing out the Steel beta. We really look forward to more of your feedback and continuing to build with this awesome, curious, and supportive community. Happy Holidays ☃️