Skyvern Review 2026

Key takeaways

• Skyvern replaces selector-based Playwright/Selenium scripts with LLM vision and reasoning, so automations survive UI changes that would break traditional scrapers.
• The open-source core (AGPL-3.0, ~23K GitHub stars) is genuinely self-hostable with bring-your-own LLM support, which is rare in this category.
• Real-world strengths are the unglamorous workflows: invoice downloads behind 2FA portals, form filling at scale, document processing.
• Cloud credits are metered and can burn fast on complex tasks; community reports describe slow steps and occasional brittle failures requiring reruns.
• Best fit: developers replacing brittle scripts, and ops teams automating portal work at scale. Not a fit if you want a lightweight, zero-infrastructure agent framework.

What Skyvern does

Skyvern automates browser workflows using AI instead of CSS selectors or XPath. You describe a task in natural language, and the agent navigates the page visually, the way a human would. The team describes the split as a Planner (holds the high-level goal), an Actor (executes the immediate step), and a Validator (checks the post-action screen state and triggers retries). That architecture matters. Single-loop agents tend to hallucinate their way through multi-step tasks; separating planning from execution with a validation pass reduces those failures.

The practical pitch is unglamorous and honest: your team logs into dozens of vendor portals to download invoices, copy-paste data, and fill forms. Skyvern automates that. It handles CAPTCHAs natively without third-party solving services, supports 2FA and TOTP authentication flows, and extracts data into whatever JSON or CSV schema you define. Every run produces an explainable summary of what the AI did and why, which is genuinely useful when a workflow fails at 2am and you need to know whether it was a selector issue or a login change.

The developer experience is the strongest part

The SDK extends Playwright directly. You get page.act(), page.extract(), page.validate(), and page.agent.run_task() layered onto the standard Playwright page object, with selector fallback when you need determinism. If you already know Playwright, the learning curve is close to flat. Trigger workflows via the Python or TypeScript SDKs, REST API, webhooks, or MCP. The MCP support is well done: Skyvern works as a tool for Claude, GPT, Gemini, or any MCP-compatible model, and it runs as a native OpenAI Apps SDK app.

Self-hosting is where Skyvern separates itself from most competitors. pip install skyvern && skyvern init, Docker Compose, an official Helm chart, and Google Cloud deployment options. Bring your own LLM: OpenAI, Anthropic, Gemini, Ollama for local models, Azure, Bedrock, OpenRouter, or any OpenAI-compatible API. If your data can't leave your infrastructure, that's a real option here, not a marketing checkbox.

One caveat on the open-source angle: the repo is AGPL-3.0, and the anti-bot and CAPTCHA-solving features are reserved for the managed cloud. Self-hosted means self-hosting the reasoning, not the full evasion stack.

Building workflows without code

Non-developers get several entry points. Copilot Chat turns plain-English instructions into a structured workflow. You can upload SOPs and have them converted into workflows, record the browser once and replay forever, or use the drag-and-drop Visual Builder. In practice, natural-language-only building works well for simple linear tasks and gets shakier on branching logic and error handling, so plan on some iteration. The Workflow Copilot v2 with self-healed retry visualization (shipped April-May 2026) helps here.

Pricing

• Free: $0, 5,000 credits (~170 actions), 1 concurrent run, basic CAPTCHA solver, datacenter proxy, country geo-targeting. Note: free credits are a one-time grant, not a monthly refill.
• Hobby: $29/mo, 30,000 credits (~1,200 actions), 10 concurrent runs, stored credentials.
• Pro: $149/mo, 150,000 credits (~6,200 actions), 25 concurrent runs, advanced CAPTCHA solver, residential proxies, city-level geo-targeting, 2FA/TOTP support, 1Password integration, team workspaces.
• Enterprise: custom credits, 100 concurrent runs, HIPAA compliance, SOC 2 report, dedicated Slack channel, Azure Key Vault, Bitwarden, custom code blocks, human-in-the-loop.

The credit math deserves attention. An "action" is roughly 30 credits, and complex pages with heavy DOM or iframes burn more. A realistic invoice-download workflow across 20 vendor portals could eat a meaningful chunk of the Pro tier monthly. The 4-hour cap on hosted browser sessions is another limit for long-running jobs, though self-hosted deployments have an uncapped runtime option.

Strengths

• Selector-free automation that survives UI redesigns, which is the core value proposition and it holds up.
• Native CAPTCHA and 2FA handling, including stored credentials with 1Password, Azure Key Vault, and Bitwarden integrations on higher tiers.
• Genuine self-hosting with bring-your-own LLM, including local models via Ollama.
• Strong integration surface: Zapier, Make, n8n, Clay, Workato, Activepieces, Pabbly Connect, plus REST API and SDKs.
• Explainable runs with per-step reasoning and action summaries.
• Active development: the June 2026 changelog alone added DeepSeek, Gemini 3.5 Flash, GPT-5-mini, and Claude Opus 4.8 as supported models, plus Google Cloud self-hosting.
• The team publishes its own benchmark (Web Bench, 5,750 tasks across the top 1,000 websites) rather than hiding behind vague reliability claims.

Limitations

• Speed and reliability on the cloud offering are the most common complaints. A Reddit user summarized it bluntly: "each step in skyvern cloud is super slow and a bit brittle, i.e. fails for a random reason and I need to rerun." LLM-in-the-loop automation is inherently slower than compiled scripts; the Route Memorization engine (compile a working path into fast Playwright code, recompile only when the site changes) is the team's answer, but it's still maturing.
• No G2 or Capterra presence as of this research. Public sentiment is scattered across Reddit and YouTube rather than aggregated review platforms, so you're doing your own diligence.
• AGPL-3.0 licensing may be a non-starter for some commercial integrations; consult your legal team.
• The free tier's 5,000 credits are a one-time grant, which the pricing page's "$0/month" framing doesn't make obvious.
• A Skyvern team member admitted on Reddit that it's "closer to an RPA-like tool than browser-use," which is honest positioning: if you want a lightweight agent framework, this is the heavier, compliance-and-infrastructure-flavored option.
• Funding details are murky: the company announced a $2.7M seed, while a LinkedIn post references $4.2M led by FCVC. Not disqualifying, but worth noting for procurement diligence.

Who it's for

Developers replacing brittle scripts: if you maintain Selenium or Playwright suites that break on every UI change, Skyvern's vision-based approach plus selector fallback is a direct upgrade path. Start on Free, expect to land on Pro.

AP and finance ops teams: the invoice-download use case is Skyvern's best story. Hundreds of vendor portals, 2FA, CAPTCHAs, no missed PDFs, structured output piped to your ERP.

Growth and lead-gen teams: Clay integration plus form filling at scale, with residential proxies and zip-code-level geo-targeting for directory work.

Regulated organizations: HIPAA compliance and SOC 2 on Enterprise, plus self-hosting with local LLMs for healthcare and finance data.

Who should skip it

If you need a lightweight, prompt-only agent framework for exploratory tasks, Skyvern's infrastructure and compliance overhead is the wrong trade. If your workflows are simple enough that compiled Playwright scripts work fine and never break, you don't need an LLM in the loop. And if you're evaluating purely on third-party review data, there isn't any aggregated yet.

Bottom line

Skyvern is one of the more credible entries in AI browser automation because it's open source, self-hostable, and honest about what it is: an RPA replacement for the era where selectors break. The developer experience is excellent, the CAPTCHA and 2FA handling solves the problems that actually kill production automations, and the explainability per run is a real differentiator. The costs are speed, credit consumption on complex tasks, and cloud reliability that still needs polish. For teams doing high-volume portal work behind logins, it's worth the free 5,000 credits to test a real workflow before committing. For everyone else, wait for Route Memorization to mature or stick with compiled scripts.