Browserbase
Give your agents access to the whole web.
5.0•12 reviews•1.8K followers
Give your agents access to the whole web.
5.0•12 reviews•1.8K 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
Browserbase
Hey Product Hunt 👋
I'm Shrey,
Over the last year, we've watched AI agents get remarkably good at using browsers. But we've also noticed something strange: every time an agent visits a website, it starts from zero.
It re-explores the interface, re-discovers buttons, re-learns navigation paths, and re-finds the same workflows it already completed yesterday.
Humans don't work that way.
Once you learn how to search Zillow listings, review a GitHub PR, or book a campsite on Recreation.gov, you don't relearn the entire website every time you come back.
Agents shouldn't have to either.
That's why we built Browse.sh, an open catalog of browser skills that agents can install and reuse across the web. Instead of exploring a website from scratch, agents load the relevant skill and execute against a known workflow.
The result is faster execution, lower token costs, more reliable outcomes, and better multi-site workflows. Today, the catalog includes 250+ skills across real websites and applications, including partner skills like submitting reimbursements on Ramp, creating projects on Lovable, extracting document data on Reducto, and many more.
And when a skill doesn't exist yet, Browse.sh can create one.
Behind the scenes, Browse.sh is powered by Autobrowse, our system that runs tasks in real browsers, analyzes traces, DOM changes, network activity, screenshots, and failures, then continuously improves the workflow until it converges on a durable strategy.
Over time, one successful browser run becomes a reusable skill that anyone can install.
Browse.sh is open source, free to use, and available today.
We'd love your feedback:
- Which websites do your agents struggle with most today?
- What skills should we add next?
- What workflows are you automating with AI agents?
We'll be around all day answering questions. Thanks for checking us out 🤞
@shrey150 This is honestly one of the more interesting takes on browser agents I’ve seen lately.
The idea that agents repeatedly “forget” workflows every session feels so obvious once you say it out loud. Humans build muscle memory for interfaces naturally, but most agents today basically wake up with browser amnesia every morning !
A few things I’m really curious about:
How durable are these skills when websites change their UI or workflows slightly? Is Autobrowse optimizing more around visual understanding, DOM structure, behavioral patterns, or a mix of all three?
How do you think about “skill drift” over time? For example, if an agent learns a workflow extremely efficiently, is there a risk it becomes too rigid and fragile when a site evolves?
The open catalog direction is super interesting too. Do you see Browse.sh eventually becoming something like an “npm for browser skills,” where agents dynamically install and compose workflows across the web?
And honestly, one of the most exciting parts for me is the multi-site workflow angle. It feels like that is where agents start becoming genuinely useful.
Really cool work. Excited to see where this goes.
Browserbase
@md_khayruzzaman thanks for the kind words! to answer your questions:
Browse.sh skills mainly describe how to accomplish a task on the current site based on the DOM structure & any internal APIs / network requests it can discover, so there's a possibility that the optimal approach goes stale or changes over time. However, each skill usually details multiple approaches & fallbacks to avoid this issue.
Skill drift is definitely a possibility, but the benefits of operating at the agent instruction level in markdown is that agents can quickly steer to different approaches, unlike scripting which either opaquely succeeds or fails. We also routinely test any skills submitted to Browse.sh to ensure that they're up to date and still detail the optimal approach!
Definitely! Thanks to the community contributing new skills, it's super easy to create "skill bundles" which compose functionality from individual sites together. You can try this yourself by telling your agent to use `browse skills find [query]` repeatedly when creating a skill bundle. In our video above, you can see us doing this to plan out a camping trip and generating a new skill bundle, `/hiking-trip-book`. The possibilities are endless!
@shrey150 this is a context window + missing memory saving loop issue, easy fix
@shrey150 Will skills have versioning, ratings, verified maintainers, or test-run results so users know which workflows are reliable before installing them?
RiteKit Company Logo API
@shrey150 The problem you're solving — agents relearning the same workflows from scratch — maps directly onto a monitoring gap that founders often hit once their product launches: knowing in real time when developers, investors, or potential partners are discussing Browse.sh across Reddit threads, YouTube tutorials, or niche dev forums. MentionFox's lead-detection engine surfaces text and audio mentions of your product from sources most tools miss, so you can catch integration requests or competitor comparisons the moment they appear, not days later. Check it out at https://mentionfox.com — in 30 seconds you can set up a Browse.sh alert and start turning organic conversations into qualified pipeline.
Product Hunt
Hey@curiouskitty !
The trigger where using a Browse.sh skill is the clear winner for an automation is if you want the fastest/simplest path to automation. You can give OpenClaw/other agents access to Browse.sh and they'll be able to run every skill/automation on the platform.
Writing Playwright scripts is completely deterministic, but takes maintenance and time to create. Generic full agent loops are usually too loosely defined and won't stay within the bounds of the task. (skills here help you save tokens as well)
Things that should NOT yet be turned into skills are workflows that need 100% determinism. We're converging to agents becoming more reliable, but there's always a chance that it goes off course.
this is very smart. i worked at Strawberry Browser before and this was a thing i thought about a lot
curious how you handle quality for open skills? like preventing stale, brittle, or low-quality workflows from spreading through the catalog
@joel_edholm We hand review all skills and requests to make sure that we're not getting straight slop into the official list.
Browserbase
@joel_edholm right now all skill runs are vetted by hand for success. We also routinely re-run skills every week to ensure quality. In the future as skill generation scales up, we may implement a consensus mechanism between different skill submissions to ensure most up-to-date info!
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_dev We do re-run skills to validate if they're still working and regenerate them upon DOM changes etc.
How much of an issue have you found captchas to be? Are websites trying to restrict agents or are they open to agentic browsing?
@rdalvi 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_pariyar 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_mallik1 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.