Skip to content
Home

/

Glossary

/

Core

/

Skyvern

Core

5 min read

Also: Skyvern AI, Skyvern Cloud, Skyvern-AI/skyvern, Skyvern open source, Skyvern browser agent

Skyvern

Quick answer

Skyvern is an open-source (AGPL-3.0) AI agent that automates browser workflows by combining Playwright with language and vision models, so it can fill forms on sites it has not seen before. You can self-host it or pay for Skyvern Cloud; for a hosted, no-code comparison, see Autonoly vs Skyvern.

Skyvern is an open-source (AGPL-3.0) AI browser automation tool from a Y Combinator S23 company that uses LLMs and computer vision to complete web tasks and workflows without hand-written selectors, available self-hosted or as Skyvern Cloud.

Last verified

What people do with Skyvern

Key facts

Made bySkyvern (Y Combinator S23)
LicenseAGPL-3.0 (open-source core)
Browser engineChromium via Playwright
ApproachVision-capable LLM reads screenshots plus page elements
DeploymentSelf-hosted (Docker) or Skyvern Cloud
Pricing modelOpen source free; Skyvern Cloud is credit-based plans

What is Skyvern?

Skyvern is an open-source tool for automating browser-based workflows with large language models and computer vision. Instead of a script that clicks #submit-btn, you describe the goal ("log in, download last month's invoice, return the total") and Skyvern works out which elements on the page to use by reading a screenshot and the page structure together. It is made by Skyvern, a company from Y Combinator's Summer 2023 batch, and is offered both as a self-hostable open-source project and as Skyvern Cloud, a hosted service.

The open-source code is licensed under AGPL-3.0. That is a copyleft license: you can use it commercially, but if you modify Skyvern and offer it to others as a network service, you have to publish your changes. The company states that the core logic is in the open-source repository, while some anti-bot measures are only in the managed cloud.

How Skyvern works

Skyvern runs a real browser through Playwright and layers an agent on top:

  • Perception. At each step it takes a screenshot and extracts the interactive elements from the page, then asks a vision-capable LLM to map the goal onto those elements. Because it reasons about what a field *means* ("this is the postcode box") rather than its selector, the same task can often run on sites with different layouts.
  • Planning and action. The model chooses an action, Skyvern executes it in the browser, and the loop repeats until the goal is met or the run fails.
  • Workflows. Beyond single tasks, Skyvern has a workflow builder that chains blocks such as navigation, extraction, file download, loops and validation, so a multi-stage process can be defined once and re-run with different inputs.
  • Code generation. Newer versions can record the actions the AI takes during a run and generate reusable code from them, so repeat runs depend less on calling the model every step.
  • It exposes an API, Python and TypeScript SDKs, and can be run with your own LLM keys when self-hosted.

    What people use it for

    Skyvern's own examples lean toward repetitive form-heavy work: requesting quotes from insurance portals, filling job applications, downloading invoices from many vendor portals, entering data into government sites and procurement systems. The common thread is a task that is the same in intent across dozens of websites that each look different, where writing and maintaining one script per site is the expensive part.

    Common problems

  • Cost per run. Every step calls a vision model. Skyvern Cloud is sold on a credit-based plan model, and the unit has changed over time (it previously charged per step); check the current pricing page before estimating volume.
  • Speed. Screenshot-plus-LLM steps are slower than a script. For high-volume jobs, generated code or a conventional script is faster.
  • Self-hosting effort. The open-source version needs Docker, a database and LLM API keys, and you take on browser infrastructure, scaling and debugging.
  • Bot detection. A self-hosted Skyvern runs a normal automated Chromium, which systems like Cloudflare Turnstile and DataDome often challenge. The stronger anti-bot handling is a cloud feature.
  • License fit. AGPL-3.0 is a problem for some companies' legal policies, especially if Skyvern would be embedded in a hosted product.
  • How it relates to automation

    Skyvern belongs to the same generation as Browser Use: agents that replace hand-written selectors with model reasoning. Where Browser Use is mainly a library you embed, Skyvern is closer to a product, with a UI, workflows and a hosted offering. Both trade the determinism of Selenium or Playwright scripts for resilience to layout changes, and both work best when the AI's run is turned into something repeatable.

    Skyvern and Autonoly

    Autonoly overlaps with Skyvern on the core idea of an AI agent operating a real browser, and differs on how you work with it. Autonoly is hosted only: you give its AI agent a task in chat, watch it in a live cloud Chromium (Playwright with patchright), and take over when it needs a human. A session can be converted into a workflow in the visual builder. Autonoly's saved workflows currently have no loop blocks and cannot sign in with stored credentials, and it offers no proxies; Skyvern's workflows do have loops and it can be self-hosted. A fuller comparison is in Autonoly vs Skyvern.

    Why It Matters

    Skyvern is one of the best-known attempts to make browser automation survive layout changes by letting a model read the page like a person. Understanding its license, cost per step and self-hosting requirements helps you judge whether it fits a task better than a script or a hosted agent.

    How Autonoly Solves It

    Autonoly is a hosted alternative: its AI agent works in a live cloud Chromium browser you can watch and take over, and a working session can be converted into a visual workflow that runs on a schedule. Unlike Skyvern it is not self-hostable, its saved workflows have no loop blocks today, and it does not provide proxies.

    Learn more

    Examples

    • An insurance broker uses Skyvern workflows to request the same quote from several carrier portals that each use a different form layout

    • A finance team self-hosts Skyvern with Docker to download monthly invoices from a list of vendor portals

    • A developer calls the Skyvern Cloud API from a backend job to fill a government registration form with data from their CRM

    Frequently Asked Questions

    Yes. The core is published on GitHub under the AGPL-3.0 license. The company says some anti-bot measures are only available in its managed Skyvern Cloud, so the self-hosted and hosted versions are not identical.

    AGPL-3.0. You can use it commercially, but if you modify Skyvern and let others use it over a network, you must make your modified source available. Some companies' legal policies restrict AGPL software, so check before embedding it in a product.

    Self-hosting is free apart from your servers and LLM API usage. Skyvern Cloud uses credit-based plans with a free tier, paid monthly tiers and enterprise pricing. The pricing unit has changed over time, so check skyvern.com/pricing for current figures.

    Yes. Skyvern drives Chromium through Playwright and adds an LLM and computer-vision layer that decides which elements to interact with, so you describe goals instead of writing selectors.

    Both are open-source AI browser agents. Browser Use is an MIT-licensed Python library you embed in your own code. Skyvern is AGPL-3.0, comes with a UI and a workflow builder, and is aimed more at repeatable business workflows such as form filling across many portals.

    The company positions CAPTCHA and anti-bot handling as part of Skyvern Cloud rather than the open-source core. A self-hosted instance runs an ordinary automated browser and will be challenged like any other. Check the site's terms before automating it.

    Related terms, automations and guides

    Where this concept shows up in practice.

    DefinitionAI AgentAn AI agent is an autonomous software system that uses large language models to perceive its environment, make decisions, and take actions to accomplish goals with minimal human direction.DefinitionBrowser UseBrowser Use is an MIT-licensed open-source Python library, made by the YC-backed startup Browser Use, that lets a large language model control a Chromium browser through a loop of observing the page and choosing clicks, typing and navigation.DefinitionPlaywrightPlaywright is Microsoft's open-source browser automation framework that provides a single API to control Chromium, Firefox, and WebKit browsers. It supports headless and headed modes, auto-waiting, and network interception, making it a leading choice for testing and web scraping.DefinitionIntelligent AutomationIntelligent automation (IA) combines artificial intelligence, robotic process automation, and process analytics to automate complex business processes that require decision-making, pattern recognition, and adaptation beyond what rule-based systems can handle.DefinitionCAPTCHACAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) is a challenge-response mechanism designed to distinguish human users from automated bots on websites.DefinitionCloudflare TurnstileCloudflare Turnstile is Cloudflare's free CAPTCHA replacement: a widget that runs invisible browser checks, issues a single-use token, and lets the site verify that token server-side through the Siteverify API.GuideAutonoly vs Skyvern: AI Browser Automation for Non-DevelopersA balanced, in-depth comparison of Autonoly and Skyvern for AI-powered browser automation. We cover architecture, features, pricing, and use cases — with honest assessments of where each tool wins.GuideBrowser Use Alternatives: Best No-Code Options for Non-DevelopersBrowser Use is a powerful open-source Python library for AI browser automation, but it requires coding skills. This guide covers the best no-code alternatives for non-developers who want the same AI-powered browser automation capabilities without writing a single line of Python.

    Stop reading about automation.

    Start automating.

    Describe what you need in plain English. Autonoly's AI agent builds and runs the automation for you — no code required.

    See Features