"Skyvern vs" is really three comparisons. One axis decides all of them.

Three different brackets, and people conflate them. The first is other browser agents, Browser Use and Browserbase, which share Skyvern's architecture (an LLM driving a Chromium tab) and differ on license, hosting, and whether they read the page with computer vision or the DOM. The second is classic enterprise RPA, UiPath, Automation Anywhere, and Power Automate, which Skyvern positions against on price and on surviving UI redesigns. The third is desktop accessibility agents like Mediar, which bind to the OS accessibility tree instead of the browser. The mistake every comparison list makes is treating these as one ranked leaderboard. They are three different surfaces, and the right pick depends on where your workflow actually runs.

They trade strengths. Skyvern reads the rendered web page with computer vision and an LLM, so it re-finds a moved button instead of breaking on a changed selector, and its usage-based pricing (free monthly credits, then historically around $0.05 per automation step) is far cheaper to start than UiPath's enterprise licensing, which runs into five and six figures a year. UiPath's advantage is reach: it was built for the Windows desktop and can drive thick-client apps, while Skyvern by its own description cannot automate desktop apps like SAP GUI. So if your workflow is web pages that change often, Skyvern is the lighter, cheaper fit. If it lives in SAP, Oracle EBS, a banking core, or any native Windows app, Skyvern cannot reach it, and the honest comparison is UiPath versus a newer accessibility-tree agent, not UiPath versus Skyvern.

This is the closest like-for-like, because both are browser agents. Browser Use is an MIT-licensed Python library you import into your own code and drive with your own LLM keys; Skyvern is AGPL-3.0 and ships as a platform with a no-code workflow builder, a managed cloud, and computer-vision page reading. Browser Use suits developers embedding an agent in an application and wanting the most permissive license. Skyvern suits non-developers who need to record and run web workflows without writing Python. We wrote a dedicated head-to-head on this exact pairing; the short version is that the license and the shape (MIT library versus AGPL platform) are the stable differences, and both share the same ceiling at the edge of the browser window.

What surface the tool binds to, because that single fact determines what it can see and where it stops. Skyvern and the other browser agents bind to a Chromium tab through the Chrome DevTools Protocol, so their world ends at the tab. Classic RPA binds to selectors and image templates, which reach the desktop but break when the UI moves. Mediar binds to the operating system accessibility tree, which exposes both the browser content and every native window around it. Once you know the binding, the rest of any comparison (pricing, vision versus DOM, star counts) is detail. The binding is the thing that decides whether a workflow finishes or stalls on step three.

Because the recorders are watching different layers. A browser-agent recorder is built on the Chrome DevTools Protocol, which only reports events inside the browser context. Mediar's open recorder is built on the OS accessibility and input layer, so its event vocabulary spans the whole machine. The WorkflowEvent enum in the open Terminator repo has 14 variants, and three of them describe things that happen outside any single browser tab: Clipboard (a copy or paste moving data between apps), ApplicationSwitch (focus leaving one app for another), and FileOpened (a native file dialog opening). The same enum carries both Click and BrowserClick as separate variants, which tells you the design treats the browser as one surface it watches, not the boundary of its world. A CDP-based recorder structurally cannot emit those three cross-application events. You can clone github.com/mediar-ai/terminator and read crates/terminator-workflow-recorder/src/events.rs to confirm it.

When the workflow never leaves the browser. If you are automating sign-ups, scraping public sites, filling forms across modern SaaS apps, or downloading invoices from web portals, and nothing in the flow opens a file picker, jumps to Excel, or touches a desktop app, then a browser-native agent is the lighter, faster choice, and Skyvern's vision approach is well suited to messy or unfamiliar sites where there is no recording to replay yet. Mediar earns its place the moment the workflow crosses into the desktop layer, which is exactly where enterprise RPA on SAP, Oracle EBS, banking cores, and EHR systems lives. We will tell you on a call if your workflow is browser-only and a browser agent is the better tool.

They bill on three different axes. Skyvern is credit-based, with free monthly credits and no credit card to start, then usage above that (historically around $0.05 per automation step), so cost scales with how many decisions the vision agent makes. Classic RPA like UiPath bills per bot or per seat under enterprise licensing that commonly reaches five and six figures a year, plus the maintenance tax when selectors break. Mediar bills $0.75 per minute of runtime with no per-seat licensing, plus a $10,000 turn-key program fee that converts to credits. Because a recorded Mediar workflow replays with no per-step model call, you are paying for wall-clock minutes of deterministic execution rather than per decision or per seat. For a high-volume, repetitive workflow on legacy screens, those three shapes of bill behave very differently.

Yes, and during a migration some teams do exactly that. Mediar reaches browser content through the accessibility tree too, so it can run the whole workflow end to end, but there is no rule that you rip out a working browser agent on day one. A common pattern is to let the existing browser agent own the pure-web stretch and let Mediar own the steps that cross into the file picker, the desktop admin app, or the green-screen terminal, then consolidate once the desktop coverage proves out. The cheaper the painful step is to migrate first, the better.