Skyvern Review

17 block types in the visual builder. CAPTCHA solving against reCAPTCHA, hCaptcha, Cloudflare and DataDome. Credentials injected from a vault so the model never sees the password. That's a real architecture decision, not a feature list — screenshots plus DOM instead of selectors, which is the actual pitch against Playwright and Selenium scripts breaking on every site redesign.

Two things stand out. One: self-hosting under AGPL-3.0 with your own LLM provider means no cloud lock-in if the credit pricing stops making sense. Two: Pro at $149/month for 150,000 credits is the tier that actually gets you 2FA and team workspaces — Hobby at $29 doesn't.

The tradeoff is real: an LLM picking clicks step-by-step is slower and less predictable than a hard-coded script, even with code caching to skip repeat reasoning. Enterprise HIPAA and SOC 2 are quote-only. Pilot it on one portal workflow before trusting it with payer or carrier logins at scale.

I run operations, which means my risk calculus is vendor portals, not internal tooling. Skyvern's pitch — screenshot-plus-DOM navigation instead of CSS selectors — maps to the actual failure mode my team lives with: a carrier portal changes its layout and the automation silently breaks on a Friday. That's a real operational cost, and a model-driven approach that adapts to layout drift is a legitimate answer, not a demo trick.

Where I slow down is credit consumption at volume. Free gives 5,000 credits, Hobby jumps to 30,000 for $29/month, Pro to 150,000 for $149/month. Fine for a pilot; I need to know what a thousand monthly invoice-download runs actually cost before I build a headcount case around it.

The code-caching feature matters more to my P&L than anything else listed: once a workflow's first run generates Playwright code, repeat runs skip the LLM call. That's the difference between a tool I can budget and one I can't forecast.

Free gets 5,000 credits, one concurrent run. Hobby is $29/month for 30,000. Pro is $149/month for 150,000, team workspaces, 2FA credential management. That's clean tier math. What's missing is the conversion rate: credits per task vary with page complexity, CAPTCHA frequency, retry count. A 50-portal-a-day operations team can't size their plan from the pricing page alone.

Enterprise adds HIPAA, SOC 2 Type II, SSO/SAML, custom credits — no number shown, as expected at that tier. Fair to quote custom there; healthcare and insurance buyers negotiating a BAA aren't procurement-blocked by that.

ROI math is workflow-specific, not platform-wide: value depends on what a manual portal task costs you per hour versus credits burned per run. Skyvern doesn't publish that ratio. You build it in the pilot.

No contract terms, renewal windows or cancellation process are visible on the plans I could find. For a credit-metered product, that's the gap that matters more than the sticker price.

Payer and carrier portal work lives and dies on the week someone redesigns a button. A selector script breaks and sits in a ticket queue until a developer fixes it; Skyvern's screenshot-plus-DOM approach is built to survive that redesign without a code change, which is the actual operational promise here, not a speed claim.

Day-3 reality is about the 17+ block visual editor and the SOP-upload path. Uploading an existing SOP document to generate a workflow is a genuinely different onboarding motion than recording clicks or writing Python, and it means the person who knows the invoice-download procedure doesn't have to be the person who builds the automation. That's a real handoff improvement for ops teams who don't have a dev on call.

The friction shows up in exception handling day-to-day: when a run stalls mid-portal, you're reading live screenshots and model reasoning logs to figure out why, which is transparent but is still a manual triage step, not an alert that tells you what broke. Credential vaulting through Bitwarden or 1Password keeps that triage from ever exposing a password, at least.

There's no mobile story here, which is fine honestly, this is a desk tool for people building portal automations, not something you'd check from your phone. What matters is the first ten minutes, and that's going to feel different depending on who you are. If you're comfortable with Python and Playwright, page.act() and page.extract() will click fast. If you're an ops person who just wants invoices downloaded from a payer portal, you're looking at a drag-and-drop editor with 17-plus block types, and that's homework before it's magic.

The live view and run recordings are the detail I'd actually watch for daily use, being able to see the browser doing its thing in real time instead of staring at a black box log is the difference between trusting a tool and babysitting it.

Month three is where vision-based navigation earns its keep over brittle selectors, but you're trading speed for that resilience every single run.

Selector-based Playwright scripts breaking on a layout change is an old, real problem. RPA vendors have promised to fix it for a decade and mostly shipped brittleness with a nicer UI. Skyvern's pitch — screenshot plus DOM tree feeding an LLM's next click — is a specific enough architecture claim that it's falsifiable, which I give it credit for.

The HIPAA/SOC 2 Type II language on the site reads confident. But both sit behind Enterprise, which has no visible number. That's the gap between marketing copy and what you can actually verify pre-sales-call. Four execution engines (Skyvern's own, OpenAI CUA, Anthropic CUA, UI-TARS) is real optionality, not vaporware — but also means your reliability story depends on a choice you make before you've piloted anything.

AGPL-3.0 plus Docker Compose self-hosting is the most honest exit story in this category. Most 'open-source' agent tools gate the useful parts behind cloud-only features. Here the self-hosted edition looks functionally whole, credentials and all.