Skyvern

Skyvern is an open-source browser automation agent that uses computer vision and large language models to interact with websites the way a human would, without relying on fragile CSS selectors or DOM structure. It ships with a no-code workflow builder, anti-bot evasion, 2FA support, and a managed cloud offering for teams that don't want to run their own infrastructure. Backed by Y Combinator and with over 21,000 GitHub stars, Skyvern targets enterprise use cases like insurance quote retrieval, government form registration, account creation, and invoice downloading at scale. Where most browser automation tools break the moment a site redesigns its UI, Skyvern's vision-first approach means it reads pages visually rather than by memory, making it far more durable in production.

Most browser automation tools are one site redesign away from breaking. A developer adds a new CSS class, wraps a button in an extra div, or migrates to a new frontend framework, and every selector you painstakingly mapped starts returning null. You’ve seen it. You’ve fixed it at 2am. Skyvern takes a different position: instead of memorizing the DOM, it reads each page visually and decides what to do in real time, the same way a person would. That architectural choice sounds simple, but it’s what separates a browser automation library from something you can actually run in production.

Quick verdict

Skyvern is the most production-ready browser automation agent available today. It’s built for teams that need durability across site changes, not just demos. The self-hosted open source path is legitimately usable, the cloud offering adds managed anti-bot evasion and debugging tooling, and the workflow builder makes it accessible to non-developers. The main caveats are cost at scale and a general-task accuracy ceiling that’s still not at human level.

What is Skyvern, exactly?

Skyvern is an open-source AI agent that uses a combination of computer vision and large language models to automate browser-based workflows. It opened its GitHub repository in early 2024, accumulated 21,500 stars faster than almost any automation project in recent memory, and shipped a managed cloud offering alongside the self-hosted version. It’s backed by Y Combinator.

The core thesis is that traditional browser automation is fragile because it depends on implementation details: element IDs, class names, XPath expressions. Any of those can change without warning. Skyvern avoids that dependency entirely. When it lands on a page, it takes a visual snapshot, analyzes it with a vision-capable LLM, identifies what the relevant interactive elements are, and then decides what action to take. It doesn’t care whether the button has an ID attribute or whether the form uses React or Vanilla JS. It cares what the button looks like and what context surrounds it.

Technically, Skyvern runs a real Chromium browser through Playwright. It exposes four core commands: act (perform an action), extract (pull structured data from a page), validate (verify a condition), and prompt (ask the LLM a question mid-workflow). These primitives compose into multi-step workflows that can branch on conditions, loop over items, and handle authentication including TOTP codes and email verification links.

The company describes the underlying architecture as “a swarm of agents” that jointly comprehend a page, plan actions, and execute them. That’s a bit of marketing language, but the practical effect is real: Skyvern scored 64.4% on the WebBench benchmark, which puts it at the top of the pack for autonomous web task completion, and it’s rated the best-performing agent specifically on WRITE tasks, the category that covers form submissions, logins, and file downloads.

It supports OpenAI, Anthropic (Claude Opus 4.7 and Sonnet 4.6), Azure OpenAI, AWS Bedrock, Gemini 3, Ollama, and OpenRouter. You pick the model. The system prompt and reasoning layer are Skyvern’s.

The features that earn the enterprise positioning

Computer vision plus LLM reasoning

The absence of selectors is the whole product. Every time Skyvern navigates to a new URL, it looks at what’s actually rendered in the browser viewport. It identifies elements by their visual appearance and surrounding context rather than by their underlying HTML attributes. A login form is a login form because it has a username field, a password field, and a submit button, not because of its class names.

This matters most when you’re running automation across dozens of different sites that you don’t control. Insurance portals, government agency sites, supplier B2B systems: these get redesigned. They migrate to new CMSes. They A/B test different layouts. A selector-based system breaks silently or with a confusing error. A vision-based system mostly keeps working because the visual structure of a form changes less often than its implementation details.

The LLM reasoning layer handles ambiguity. When a page has multiple possible next steps, or when an element label is ambiguous, the model reasons through what the intended action should be given the workflow context. That’s meaningfully different from rule-based automation where you have to anticipate every edge case in advance.

Workflow definitions and reproducibility

Skyvern ships a visual, no-code workflow builder. You define a workflow as a graph of steps: navigate here, fill this form, extract that data, branch on this condition. The workflow stores as a structured definition you can version, share, and re-run. This is closer to how enterprise automation teams think about processes than to how developers think about scripts.

The conditional logic is genuine. Workflows can branch based on extracted values. If the site returns an error message, take path A. If the form submitted successfully, take path B. Loops let you run the same sequence over a list of inputs, which is how you’d process a batch of insurance quotes or a list of vendor portals.

The workflow definitions are durable in the same way the element detection is durable. Because you’re describing intent rather than implementation, the workflow doesn’t break when a site changes its layout. That’s the promise, at least. Reality is more nuanced: complex multi-step workflows on poorly structured sites still require tuning.

Anti-bot evasion that actually works

Running any automated browser against sites with bot protection is a game. Skyvern Cloud plays that game better than most. The vision-based interaction pattern already helps: the system doesn’t move through the DOM programmatically the way Selenium does, and its interaction timing is less mechanical. On top of that, the cloud offering layers anti-bot evasion handling that helps sessions survive more aggressive detection systems.

CAPTCHA handling is included. Skyvern handles a range of CAPTCHA types as part of the cloud service. This isn’t a guarantee of 100% success against every protection system, and sites running Cloudflare Enterprise or Akamai Bot Manager at maximum aggression will still challenge sessions. But for the typical B2B portal, government website, or commercial service that uses standard bot protection, Skyvern’s combination of human-like visual interaction and active evasion support changes what’s possible.

2FA authentication is also genuinely supported. You can configure TOTP-based 2FA, email code retrieval, and SMS code handling. It integrates with Bitwarden, 1Password, and LastPass for credential management. These are requirements for automating real production accounts, not sandboxes.

Hosted cloud for production deployments

Self-hosting Skyvern requires Python 3.11, NodeJS, and a working Docker setup. It’s not onerous for a developer, but it’s a real infrastructure responsibility: you’re running browsers at scale, managing compute, and handling failures. Skyvern Cloud removes that. You get managed infrastructure, browser sessions that spin up on demand, and a debugging interface that streams live browser viewport video so you can watch what the agent is doing in real time.

The livestream debugging is underrated. Browser automation at scale fails in weird ways. Watching a live or recorded session of a task that went wrong is fundamentally faster than parsing logs. Skyvern Cloud ships this out of the box rather than making you instrument it yourself.

The cloud also handles session management, retry logic, and the operational overhead of keeping automation running reliably. For enterprise teams that don’t have a dedicated infrastructure engineer, this is the difference between a project that ships and one that stalls.

API and SDK for orchestration

Skyvern exposes a REST API, a Python SDK, and a TypeScript SDK. This means browser automation can become a backend service called by your existing applications. You’re not running a script locally; you’re calling an API that returns results when the task completes.

This is important for the use cases Skyvern is built for. When your compliance system needs to pull a form submission confirmation from a government portal, it shouldn’t need a human in the loop. It should call an API, get a result, and continue. Skyvern’s API layer enables that architecture.

There are also native integrations with Zapier, Make.com, and N8N for teams that work in no-code automation contexts. That broadens the accessible user base beyond developers.

Pricing

Skyvern is open source under a permissive license. Self-hosting is free, and the GitHub repository gives you everything you need to run it on your own infrastructure. This is a real option for development teams with the technical capacity to manage it.

Skyvern Cloud is the managed product. Pricing starts around $99 per month for lower task volumes and scales up from there, with higher-tier plans in the $499 per month range for teams needing more capacity. Enterprise contracts are available and are the path for high-volume production deployments or organizations with compliance requirements around where their data lives and how their browser sessions are managed.

The pricing model makes sense for the value delivered at enterprise scale. Running reliable, anti-bot-aware browser automation across hundreds of workflows per day is genuinely expensive to operate, and the cloud pricing reflects that. Where it gets uncomfortable is for smaller teams or individual builders doing moderate task volumes: the jump from free self-hosted to paid cloud is meaningful.

What’s worth factoring in is the cost of the alternative. A developer maintaining a Playwright-based automation system, debugging failures, updating selectors after site changes, and handling auth edge cases is probably spending more hours per month than the Skyvern Cloud subscription costs. The ROI equation tilts toward the managed product faster than it looks at first.

There’s no public pricing calculator, and exact per-task rates for higher volumes require contacting the sales team. More pricing transparency would help prospective buyers evaluate fit before jumping on a call.

Where Skyvern wins and where it doesn’t

Skyvern wins on durability. If you’re automating workflows across sites you don’t control and you need them to keep working after site changes, Skyvern’s vision-first approach is the right architecture. It’s not invincible, but it’s materially more resilient than anything selector-based.

It wins on structured, repeatable workflows. Form submissions, account creation, data extraction, file downloads: these are Skyvern’s sweet spot. The 64.4% WebBench accuracy and top ranking on WRITE tasks aren’t accidental. The system is optimized for the kinds of tasks enterprise operations teams actually run.

It wins on authentication complexity. Handling real-world login flows with 2FA, password managers, and session management is table stakes for production use. Most browser automation tools treat auth as an afterthought. Skyvern treats it as a first-class feature.

Where Skyvern struggles: fully open-ended, exploratory tasks where the right action isn’t inferrable from the page context alone. It’s not an autonomous research agent. It’s an automation agent. The distinction matters. Give it a clear goal and a repeatable process, and it performs well. Ask it to make judgment calls that require outside-the-page context, and the accuracy drops.

Complex multi-step workflows on badly designed sites, ones with ambiguous UI, overlapping modals, or inconsistent navigation, also require more tuning than the docs suggest. Expect iteration.

Who Skyvern is built for

Skyvern fits three categories of user, and it fits each of them differently.

Enterprise operations teams are the primary audience. If your company submits forms to government portals, pulls data from supplier systems, opens accounts across insurance or financial services providers, or manages any high-volume repetitive web workflow, Skyvern is designed for you. The workflow builder, the cloud infrastructure, the API layer, and the anti-bot handling are all optimized for this use case.

Developers building products are the secondary audience. If you’re shipping a product where browser automation is a backend service, the API and SDK make Skyvern a legitimate component. You don’t have to solve the vision-based automation problem yourself; you call an API and handle the result.

Technical teams that want control but don’t want to build from scratch benefit from the self-hosted path. The open source base is mature enough to customize, and the community around it is active.

Solo builders and individual developers doing low-volume, one-off automation tasks are probably not the core audience. The overhead of either self-hosting or paying for cloud is hard to justify at that scale.

Skyvern vs the alternatives

Browser Use is the most frequent comparison, and the distinction is real. Browser Use is a Python library. You import it, write code, and control every aspect of how it calls the LLM and how it interacts with the browser. It’s flexible and gives developers fine-grained control. Skyvern is a product with managed infrastructure, a visual UI, a cloud offering, and an opinionated workflow model. Browser Use is faster to prototype with. Skyvern is more reliable to operate at scale. If your team includes non-developers who need to build or modify workflows, Skyvern wins immediately.

MultiOn takes a more consumer-facing approach. It’s built around the idea of a personal browser agent you direct conversationally. Skyvern is built around the idea of defined, reproducible workflows you run programmatically. They’re targeting different problems. MultiOn is better for interactive, one-off personal tasks. Skyvern is better for automated, high-volume business workflows.

OpenAI Operator is the closest product-level competitor. Operator is a managed, cloud-only browser agent from OpenAI, and it’s capable. The key differences are model lock-in (Operator runs on OpenAI models only; Skyvern lets you choose), open source availability (Skyvern has a self-hosted path; Operator doesn’t), and workflow tooling maturity. Operator is more polished for consumer and lighter business use cases. Skyvern’s workflow builder and API layer are currently more suited to enterprise automation that needs programmable, reproducible process definitions.

For teams already in the Anthropic ecosystem using Claude Opus 4.7 or Sonnet 4.6, Skyvern is particularly interesting because you can run it against Anthropic’s models and stay within your existing model contracts. Anthropic Computer Use is the underlying capability, but Skyvern wraps it in production infrastructure that Anthropic’s raw API doesn’t provide.

Getting started

The fastest path is Skyvern Cloud. Sign up, get API credentials, and you can run your first task with a single API call or through the web dashboard. No infrastructure to configure.

For self-hosting, you’ll need Python 3.11 and NodeJS. Clone the repository, copy the sample environment file, add your LLM API keys, and run the Docker Compose setup. The documentation walks through the process step by step. Expect to spend thirty to sixty minutes getting a working local installation, plus time tuning your LLM model choice and API rate limits for your specific use cases.

The workflow builder is the right starting point for most enterprise users. Define your workflow visually, test it against a target site, and then export or trigger it via API when it’s working. Iteration is faster in the UI than in code.

For developers integrating Skyvern as a backend service, the Python SDK is the cleanest entry point. The API reference is thorough and the SDK covers the full surface area. Zapier and Make.com integrations are available for teams that want to connect Skyvern to existing automation stacks without writing code.

The community is active and the GitHub issues tracker is responsive. That matters when you hit the inevitable edge case on a protected site.

The bottom line

Skyvern is what browser automation looks like when you build for production from the start. The vision-first architecture is the right call for workflows that need to survive the real web, where sites change, bot protection is real, and auth flows are complicated. The managed cloud offering makes it accessible to teams without dedicated infrastructure engineers, and the open source base keeps self-hosting viable for teams that need it.

It’s not perfect. Accuracy on general tasks still leaves headroom, pricing transparency could be better, and complex workflows on messy sites require real effort to tune. But for enterprise teams automating high-volume, structured web workflows, there’s nothing else that checks as many boxes right now. If you’re evaluating browser automation for production, Skyvern should be on your shortlist.

Key features

• Computer vision plus LLM reasoning for element detection without brittle selectors
• Visual workflow builder with conditional logic and multi-step branching
• Anti-bot evasion and CAPTCHA handling for protected sites
• Hosted cloud with managed infrastructure and debugging livestream
• REST API, Python SDK, and TypeScript SDK for programmatic orchestration
• 2FA support including TOTP, email, and SMS codes
• Password manager integrations with Bitwarden, 1Password, and LastPass

Pros and cons

Who is Skyvern for?

• Enterprise compliance teams automating government form submissions and regulatory filings
• Insurance and fintech companies pulling quotes or opening accounts across dozens of provider portals
• Operations teams automating invoice downloading and procurement workflows across supplier sites
• Developers building products that need reliable, API-driven web automation as a backend service

Alternatives to Skyvern

If Skyvern isn't quite the right fit, the closest alternatives are browser-use , multion , and openai-operator . See our full Skyvern alternatives page for side-by-side comparisons.