The Skyvern journal

Another day another surprise! A short one today: Web bench is live on Product Hunt. We would love your support here: https://www.producthunt.com/products/web-bench

TL;DR One toggle → your automations run on-device, no cloud hop. # .env ENABLE_OLLAMA=true LLM_KEY=OLLAMA OLLAMA_MODEL=qwen2.5:7b-instruct OLLAMA_SERVER_URL=http://host.docker.internal:11434 Why you’ll care * Zero data leaves the box – ideal for air-gapped or PII-heavy flows, * Drop-in swap – test the latest gemma model or qwen model with a one line change * Run custom fine-tunes - fine-tune your own open source LLM and try it out with Skyvern

Today we’re releasing two small Python packages that let you trigger Skyvern browser tasks directly from LangChain or LlamaIndex code. * skyvern-langchain – LangChain Tool + Agent helpers * skyvern-llamaindex – LlamaIndex Tool + Runnable helpers Both follow the same pattern: create a task, wait for it if you like, or hand it off to Skyvern Cloud. Nothing else to set up. Installation pip install skyvern-langchain # LangChain pip install skyvern-llamaindex # LlamaIndex

TL;DR Plug any OpenRouter-hosted model into your automations with one env toggle. No code changes, no redeploys. # docker-compose.env ENABLE_OPENROUTER=true LLM_KEY=OPENROUTER OPENROUTER_API_KEY=sk-… OPENROUTER_MODEL=qwen/qwen2.5-vl-32b-instruct Why you might care * BYOM (Bring your own model) – swap models in minutes. * Full Skyvern feature set – async tasks, vision support, streaming, same API. Under the hood LLMConfigRegistry.register_config( "OPENROUTER", LLMConfig(

Today we’re launching a major super-power for your browser automations: securely add credit cards once and let Skyvern handle checkouts end-to-end Why this matters * One-click setup – Save a card in seconds; via API or via our UI. Skyvern stores it inside a secure vault. * Hands-free automations – Your agents can now pay for subscriptions, place test orders, validate carts, or top-up balances while you sleep. * Enterprise-grade security – Card numbers are never exposed to your workflows; o

Most automation tools are closed boxes. Skyvern is different. It’s one of the only open-source, browser-native AI automation platforms — and it’s built for developers who want flexibility, control, and infrastructure-grade power. Here’s what you should know if you're evaluating Skyvern from a dev perspective 👇 🧠 LLM Flexibility Built In Skyvern supports multiple LLM providers out of the box, including: → OpenAI (GPT-4o, GPT-4 Turbo, GPT-4o-mini) → Claude 3 (Haiku, Sonnet, Opus, and Claude

Every agent vendor right now claims state-of-the-art performance on browser tasks. Skyvern, Browser-use, OpenAI's Operator are all shipping numbers that look strong on paper. But the benchmark most teams use to compare these agents, WebVoyager, covers just 643 tasks across 15 websites. That's a narrow slice of the internet, and it mostly tests agents on read-heavy tasks where the real friction hasn't kicked in yet. The harder problems are logging in, solving 2FA, filling out forms, and downloadi

If you're still copying data from one tab to another or downloading reports by hand, you're giving away hours you'll never get back. Modern AI browser agents, like the open-source platform Skyvern, let you hand off those chores without writing a single line of code. Browser automation is the mechanism here, and it sits inside a broader category called Agentic Process Automation (APA), where autonomous multi-step operation and exception handling are the actual product. This article explains, in p

Here's O3's definition: Nirvana fallacy: rejecting a workable solution because it isn’t perfect—an easy way to stall real progress while chasing an imaginary ideal. This is a common pitfall that engineers like me run into "But it won't work in this edge case" "But did you consider this situation that's only going to occur if 5 other things happen in a row?" This line if thinking is important when you're designing system for scale If you're working at Google with 1 billion users, even a 0.1

Some business-critical chores still hide behind login screens, pop-ups, and "Download" buttons that no API can reach. Skyvern's AI browser agent bridges that gap. Give it a plain-English description of the job, and it sees the page like a human, decides where to click or type, and gets the work done. No scripting, no brittle screen recordings. TLDR: * Browser-based workflows hide behind login screens and portals that have no API, making scripted tools fail silently when layouts change. * Fai

From virtual assistants to productivity SaaS, the next generation of AI companies needs a rock-solid way to interact with the web the same way humans do. Skyvern gives you that foundation: an open-source, prompt-based browser agent that combines large-language models with computer vision to click, type and scroll its way through any site you throw at it—no brittle XPaths or head-scratching scripts required.  This post walks through the why and how of building a profitable AI venture on top of t

⭐ We hit 13K GITHUB STARS!!! What a big milestone!! But.. what if we told you that we stopped tracking star count as a top line metric 6 months ago? Star count is similar to user signups. It's a vanity metric. It's a great metric that's guaranteed to be up-and-to-the-right, and help you feel good, but you can't eat Github stars for dinner It's a representation of "this is interesting and cool!" which doesn't always translate well to "this will become a sustainable business" What are better

I've seen a lot of skepticism around MCP on the internet. I'm here to convince you everyone is wrong MCP is real and valuable. It's an extremely simple protocol – it's an API spec that tells an LLM how to use a tool. Why is this valuable? Tools for LLMs are the same as tools for us. You can use a hammer to put a nail into a piece of wood. You can use a browser to browse the web. It's the same for LLMs. Tools help them do things they weren't able to do before So.. why do we need MCPs? Why not

Learn about the tips and tricks the Skyvern team uses to launch new features across many different social media channels

We're building Skyvern, an open-source Agentic Process Automation (APA) platform. It uses computer vision and AI reasoning to read web pages visually, work through multi-step authentication flows, and complete browser-side tasks without brittle selectors or code that breaks the moment a site changes. Browser automation is the execution layer; the APA platform handles the broader job of autonomous, multi-step operation across portals and workflows that have no direct API to call. Over the weeken

Got invited to a private YC Open Source conferences last week Here are my top 10 takeaways from it! 1 - The old age of open source (pre-2015) was crazy 2 - Apparently it was normal to donate your open source project to the Apache organization? 3 - There are big differences between Apache 2.0 and GPL-3 as a license. 4 - Cloud providers stealing your project is largely a myth (see Redis + Azure partnership). GPL-3 Makes it impossible for this to happen 5 - OSI approved licenses matter to the com

Gather ‘round! Our integration in n8n (community node) is finally here! n8n is a workflow automation platform that empowers technical teams to build complex, multi-step workflows using their 400+ native integrations, with an even richer community node ecosystem Why Integrate into n8n? Is it fate? Is it science? Don’t know, but there has been a lot of overlap between Skyvern users and n8n users. So, we’ve made an integration that connects the two more easily. Here’s what we do know: people o

You've probably dealt with those outdated websites or really clunky online portals and  you know how frustrating they can be. Maybe you’ve had to log into a government site with a somewhat ancient interface, submit a job application through a very slow, multi-step process, or extract data from a vendor portal that feels like it hasn’t been updated in decades. For anybody or business that typically relies on these platforms for important work, like retrieving invoices; using regular automation

How do you know you built something people wanted? When you get users start using it and complaining before you even have a chance to announce its launch This has happened at Skyvern many times... but the most recent is the most fascinating We secretly launched support for credentials almost 2 weeks ago. As soon as the deploy went live, we got a message: "Hey this isn't working". We had a ton of bugs but we figured we could fix them before people noticed they were there.. 1. The error mess

🗣️ You can now store credentials directly in Skyvern! Until now, the only way to manage credentials was through Bitwarden. Time and time again, we heard the same feedback—signing up, onboarding, and configuring was was annoyingly high maintenance. With this release, you can now add, save, and link credentials into your workflow without leaving Skyvern. Three easy steps: add credential, add a credential parameter in workflow, and link! Whether you’re an individual looking for a quick start or

Hit a major Skyvern milestone Skyvern finished its first autonomous order. End to End. This wasn't a demonstration -- this was a real order being placed to automate something very core to a business we've been working closely with Here are all of the steps it has to perform for it to work correctly: 1. Navigate to an e-commerce website 2. Log in 3. Clear the cart if there's already products there 4. Go through a list of products 5. Check if the specified variant exists and is avail

I don't think we can deny the fact that automation has changed how we approach repetitive tasks by saving time and eliminating the monotony of manual work. However, when automating tasks in browsers, errors can quickly become somewhat of a roadblock. From stuff like page timeouts to CAPTCHA challenges, automation can get pretty frustrating if not handled effectively. There's good news though! You can easily handle errors in browser automation with some good strategies and the right mindset. Let

🌟 TWELVE THOUSAND 🌟 Here are some interesting star milestone facts 1. Going from 0 - 5K took as long as going from 5K - 12K 2. We stopped looking at star count as a top line metric when we launched Skyvern Cloud 3. Star count is not correlated with our discord user growth 4. Companies stopped caring about our star count after 5K Github stars -- 5K hit the sweet spot of credibility 5. Star count is not the same as product market fit 6. It is a great proxy for market pull + interes

In the words of Gen Z, we were cooking this month and have a lot of interesting changes to share with you! Launch Skyvern 2.0 “Observer” We dropped Skyvern 2.0! It’s a major re-design of Skyvern’s agent, introducing a planner and a validator to allow Skyvern to autonomously complete complex tasks on the web. On WebVoyager, Skyvern 2.0 scores state of the art, completing 85% of tasks prompted, beating out Google Mariner and other major competitors in the field. Check out the results at https:/

Excited to announce our latest collab! Ben builds the coolest AI agents on n8n and Relevance AI and gives them access to a bunch of tools to let them complete REAL WORK. You know what the most ambitious and underrated tool is? Giving an AI agent access to the internet. This is where Skyvern comes in. In this one, he builds a purchasing agent and contact form agent controlled by WhatsApp Voice messages with n8n. n8n handles the agent interactions, and Skyvern handles all web operations. HOW