Browse.sh - Give your agents muscle memory for automating the web
byβ’
browse.sh β 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.
Replies
DiffSense
Anyone made a ryanair skill yet? I guess we need GPT 6 or 7 before its possible to navigate that website with AI π¬
@conduit_designΒ I was navigating the TAP airlines website yesterday and was literally feeling that pain :D
The 'muscle memory' framing is apt: it's really a domain-specific replay buffer for web interactions. We've been building AI agents that automate customer workflows and session state management across sites is genuinely hard. How does Browse.sh handle sites that frequently change DOM structure? Does the agent re-learn from scratch or do partial cache invalidation?
@retain_devΒ We do re-run skills to validate if they're still working and regenerate them upon DOM changes etc.
Stripo.email
This makes a lot of sense. Teaching agents the same website every day feels a bit like deleting your browser history before every session π Congrats on the launch!
@alina_tyslenok_Β 100%, imagine trying to do your job everyday but you forget all your passwords and where all the information lives. Browse.sh gives your agents context on how it should do tasks (based on how it's been done before).
How much of an issue have you found captchas to be? Are websites trying to restrict agents or are they open to agentic browsing?
@rdalviΒ We partner with companies like Cloudflare on Web Bot Auth, ensuring that our platform enables good actors to use websites, and keeps out any bad actors with malicious intent. Here's some more information on our work on identity: https://www.browserbase.com/identity
agents have no muscle memory, and that's a real cost in tokens, in time, in reliability.
the open catalog model is interesting. curious how skill quality is managed as it scales. is there a review layer or does it mostly rely on community signal to surface what actually works?
good luck with the launch!
@riya_pariyarΒ We currently do a lot of the review manually. As we scale there will be some process automation, but we're white gloving to ensure only high quality skills enter our catalog.
Memory seems useful when an agent is actively working, but the harder problem feels like deciding what deserves to be remembered in the first place.
Have you found the biggest gains come from long-term memory across sessions, or from reducing context loss within a single workflow?
@zaid_mallik1Β It's not just memory but also a fully optimized workflow. We use Autobrowse under the hood, which just iteratively improves browser skills (similar to autoresearch from Karpathy). In some skills you'll find that the model found a hidden API endpoint, which may be more token efficient than spinning up a full browser to do a task.
@kyle_jeongΒ The hidden API endpoint example is a really interesting observation. It feels like a lot of the gains now are coming from finding more efficient execution paths rather than making models smarter.
We're building something in the AI workflow space ourselves and have been running into similar questions around context, orchestration, and cost efficiency. Planning to launch soon, so it's always useful seeing how others are approaching these problems.
fire proposition. played around with browser automation to scrape scholarships off the web; wonder if something like this could help me apply to them automatically asw ...
@sanjith_sambath1 Yep you should be able to create a skill to help you apply as well. If Browse.sh gets you a scholarship I hope we get 1% haha
The "agents start from zero every time" framing is exactly right. I drive a browser agent every day for marketing routines and the recurring tax is re-finding the same comment box, the same upvote button, the same auth modal. Reusable SKILL.md per domain is the unlock - and an open catalog turns it into a network effect. Following.
@shrey150 The idea that agents should have muscle memory like humans do is so obviously right that itβs wild nobody shipped this sooner. Watching an agent re-explore the same GitHub PR workflow for the tenth time feels like watching someone forget how to use a doorknob... and now thereβs finally a fix. Yaay!
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?