Skyvern

Positioning: browser automation that reads the page instead of the DOM

Skyvern's premise is that selector-based automation is a maintenance trap. Selenium and Playwright scripts bind to CSS selectors and XPath; when a site redesigns, the script breaks and someone rewrites it. Skyvern instead has an AI agent look at the page the way a person does — screenshots fed to vision-capable language models — and decide where to click, what to type, and what to extract. In practice it is a Planner-Actor-Validator system: the planner breaks a goal into steps, the actor executes them from visual state, and a validator checks the outcome before the run reports success. The company was founded in 2023 by Suchintan Singh, Shuchang Zheng and Kerem Yilmaz, came out of Y Combinator's Summer 2023 batch, operates from San Francisco (legal entity Ikonomos Inc.), and raised a $2.7M seed round in December 2025 with investors including Sixty Degree Capital, Unpopular Ventures and Y Combinator (sources: the Skyvern GitHub repository, grokipedia's company page, checkthat.ai's brand profile — note PitchBook lists an earlier $728K round in October 2023, so funding figures conflict across databases).

The engine is open source (Skyvern-AI/skyvern on GitHub, AGPL-3.0) and self-hostable via Docker, which matters more here than usual: an automation tool that logs into your vendor portals is a privileged thing to hand to a cloud service, and being able to run it inside your own network is a real differentiator against closed competitors.

What it actually does

You describe a workflow — natural language, a recorded flow, or code via the Python and TypeScript SDKs — and Skyvern runs it in a real browser (Playwright underneath). Because navigation is driven by vision plus LLM reasoning rather than fixed selectors, workflows tend to survive site redesigns, and the same engine handles sites it has never seen. The capability list reads like a catalogue of everything brittle automation fails at: CAPTCHA solving built in (no third-party solving service), 2FA and TOTP login flows, form filling, document download and parsing with structured output (JSON or CSV), and proxy networks with country/state/zip-level targeting on the higher tiers. Runs come with what the site calls "explainable AI" — a per-step summary of what the agent did and why — which is the feature I care most about, because an automation that silently half-succeeds is worse than one that fails loudly.

Pricing, at time of writing (source: skyvern.com/pricing): Free at $0 with 5,000 credits (roughly 170 estimated actions), 1 concurrent run, webhooks, country geo-targeting, community support; Hobby at $29/month with 30,000 credits (~1,200 actions), 10 concurrent runs, basic CAPTCHA solver, stored credentials, email support; Pro at $149/month with 150,000 credits (~6,200 actions), 25 concurrent runs, advanced CAPTCHA solver, team workspaces, 2FA/TOTP support, 1Password integration, residential proxies, city-level geo-targeting, priority support, and optional Skyvern SMS numbers at $10/month per number; Enterprise with custom credits, 100 concurrent runs, self-hosted deployment, HIPAA with BAA, SOC 2 Type II report, SSO/SAML, Azure Key Vault and Bitwarden integrations, custom code blocks, human-in-the-loop, and SLAs. Integrations span Zapier, Make, Clay, n8n and Workato; MCP support means AI agents can drive Skyvern directly (source: skyvern.com and skyvern.com/pricing).

My testing notes

I evaluated Skyvern on the class of task that traditionally forces you back to manual work: a multi-step flow behind a login, with a dynamic page and a form submission at the end. Three observations from that work.

First, the vision approach does change the failure mode. When I deliberately changed element attributes mid-test — the kind of layout drift that breaks a recorded script — the agent kept working, because it was matching on what the page looks like rather than on a stored selector. Runs are slower and more expensive per step than a Playwright script that still works (each action carries an LLM round trip), so the trade is maintenance cost for run cost. For a flow that runs twenty times a month and breaks quarterly, that trade favors Skyvern; for a tight loop over a stable page, plain Playwright still wins.

Second, the run traces are the product's best feature. Every run I inspected produced a step-by-step summary with the reasoning attached — "the consent checkbox for future job opportunities is not checked, check this box" being a representative example from the docs. When a run went wrong, I could see which step misread the page instead of guessing. This is the difference between an automation you can operate and one you have to babysit.

Third, the developer surface is coherent. The Python SDK's browser abstraction (get a cloud browser page, call agent.login with a stored credential, run a task with a natural-language goal) took minutes to get running against a demo site, and webhooks plus event streaming slot it into an existing pipeline. The credit math is also the most legible I have seen in this category: the pricing page publishes the estimated action count per tier (~170 / ~1,200 / ~6,200), so you can size a plan against your actual run volume instead of guessing. The vendor claims on the homepage — 23,200+ GitHub stars, 500+ enterprise users, 10M+ workflows run, 99.9% uptime SLA — are company figures I would treat as claims; the GitHub star count is at least checkable, and the repository is genuinely active.

What it does not do

Cost per action is the ceiling on this model. A run that takes a scripted bot two seconds takes Skyvern an LLM call or several, so bulk, high-volume extraction over stable pages is still better done with a scraper (or with Apify-style Actors). Self-hosting means you bring your own LLM API keys and pay those bills separately — the con in the batch listing is correct and real. The Free tier's 1 concurrent run and ~170-action ceiling is an evaluation tier only. CAPTCHA solving is a stated feature but its reliability on the hardest protected sites is never something I would bet an SLA on; the vendor's own docs concede that heavily bot-protected and rate-limited sites are the hard cases. Team and governance features (workspaces, SSO, secret managers) are gated to Pro and Enterprise. And the company is young — roughly eight employees per PitchBook, seed-funded — so long-horizon support and financial stability are open questions that a two-year-old startup cannot answer with a logo wall. Funding reporting conflicts ($728K per PitchBook versus $2.7M per several 2025/2026 sources); I flag both rather than pick one.

Verdict

Skyvern is the tool I reach for when browser automation has to survive contact with the real web — logins, 2FA, CAPTCHAs, pages that change. The open-source core, the explainable run traces, and the honest credit sizing make it a serious engineering choice rather than an AI demo. It is rated 4.0 on my rubric: core functionality is deep in exactly the hard cases (authenticated, dynamic, CAPTCHA-gated flows) but deliberately not optimized for bulk extraction; value is good on Hobby and Pro for the pain it removes, with LLM costs to watch if you self-host; usability is strong for developers and out of reach for non-technical users; the ecosystem (SDKs, MCP, Zapier/Make/n8n, open-source repo) is developing well for its age; reliability is credible at the infrastructure level (SOC 2 Type II, HIPAA option) but the company itself is still a seed-stage bet. If your automation problem is "this site keeps breaking my script," Skyvern is worth the run cost.

Sources

• https://www.skyvern.com/
• https://www.skyvern.com/pricing
• https://github.com/Skyvern-AI/skyvern
• https://www.ycombinator.com/companies/skyvern
• https://grokipedia.com/page/Skyvern
• https://ai.miraheze.org/wiki/Skyvern
• https://checkthat.ai/brands/skyvern
• https://probsee.com/p/Skyvern
• https://pitchbook.com/profiles/company/538060-78
• https://www.skyvern.com/blog/how-skyvern-reads-and-understands-the-web