Browserbase
Give your agents access to the whole web.
5.0•12 reviews•1.9K followers
Give your agents access to the whole web.
5.0•12 reviews•1.9K followers
Browserbase makes the web as reliable and programmable as APIs.
This is the 3rd launch from Browserbase. View more
Browse.sh
Launched this week
browse.sh — an open catalog of browser automation skills for any website. Find reusable SKILL.md recipes that teach AI agents to complete tasks online, and install them with the browse CLI.
Free
Launch Team / Built With
Agent Browser Shield
Congrats the on the launch! We've been using Browserbase to benchmark https://github.com/pixiebrix/agent-browser-shield and prototyping how to enforce/apply the browse.sh content at the harness layer instead of context layer! Love it!
@tschiller browse.sh skills in the harness will definitely give your agents the edge they need, enforcing the workflow rather than just hoping that the agent picks it up.
The "npm for browser skills" framing is exciting and honestly a little scary. With an open catalog and community contributions, when two skills exist for the same site, how does an agent decide which one to trust? And what stops a subtly wrong or even malicious skill from spreading before your weekly revalidation catches it? Feels like skill provenance and reputation become the real product once the catalog gets big.
the CLI install pattern is the right distribution mechanism for developers but the interesting question is who maintains skills for sites that actively try to block automation. linkedin and some other platforms regularly update their markup specifically to break scrapers and automation tools. does browse.sh have a position on skills for sites that have terms of service prohibiting automated access or is that left entirely to the skill author and the person installing it
Pod
This is a great use of memory! How do you deal with information recency?
@patrick_monnot A little confused by the question, but I'll attempt to answer. We revalidate the skills every week so our skills stay up to date (if the page changes or data updates, our skills should too). Memory here is like a weak cache, we invalidate if our data is stale.
There’s some level of stochasticity in the “perfect recipe” and resulting SKILL.md output for a given task. If multiple users each have a distinct generated skill to perform the same task uploaded to Browse.sh, how does the platform resolve these redundancies?
I like the idea of agents not starting from zero every time. Re learning the same website flow feels like wasted time and tokens. How do you handle skills when a website changes its layout or button names?
Is there any way to bypass bans? For example, Amazon doesn’t seem to like AI agents very much and bans them outright.