Part of our Browser Automation MCP category.

At a glance: 21,300 GitHub stars, 1,900 forks, ~4,492 commits, 11 employees, last commit April 2026, v1.0.31 (April 14, 2026), 75+ tools, AGPL-3.0 license, ~354K total PyPI downloads, PulseMCP ~353K all-time visitors (#126 globally, ~3,100 weekly). YC-backed, $2.7M raised. Founded by Suchintan Singh, Kerem Yilmaz, and Shuchang Zheng.

Skyvern takes a fundamentally different approach to browser automation than Playwright MCP or Browserbase. Where those servers rely on DOM selectors and accessibility trees, Skyvern uses computer vision and LLMs to understand what’s actually on the screen. A “Submit” button is a “Submit” button whether it’s a <button>, a styled <div>, or an image — if a human can see it, Skyvern can interact with it.

Built by Skyvern-AI and backed by Y Combinator, the platform has grown to 21,300+ GitHub stars and is used for form automation, web scraping, and RPA-adjacent workflows where traditional selector-based automation breaks constantly.

What It Does

The MCP server exposes 75+ tools organized across seven categories:

Browser Session Management

  • skyvern_browser_session_create / close / list / get / connect — open, manage, and reuse browser sessions, including connecting to a local browser via Chrome DevTools Protocol (CDP)

Browser Actions (Natural Language)

  • skyvern_act — perform actions using natural language (“click the login button”, “fill in the email field with test@example.com”)
  • skyvern_navigate — go to a URL
  • skyvern_click / type / hover / scroll / select_option / press_key / wait — granular browser control with CSS/XPath selectors or natural language fallback

Data Extraction

  • skyvern_extract — pull structured JSON from pages using natural language descriptions and optional schemas
  • skyvern_screenshot — capture page screenshots
  • skyvern_evaluate — run JavaScript in the browser context

Validation

  • skyvern_validate — assert page conditions in natural language (returns true/false with reasoning)

Credentials & Login

  • skyvern_login — automated login with credential vault integration (Bitwarden, 1Password, Azure Key Vault) and automatic 2FA/TOTP handling
  • skyvern_credential_list / get / delete — manage stored credentials

Workflows

  • skyvern_workflow_create / run / status / get / update / delete / cancel — build, run, and manage multi-step automations with 23 block types

Additional Capabilities

  • Tab management and iframe switching
  • Drag-and-drop and file upload
  • Clipboard access and network/console inspection
  • HAR recording and auth state persistence
  • Folder and script management

The key differentiator is the vision-based approach. Instead of breaking when a developer changes a button’s CSS class or moves an element, Skyvern adapts because it “sees” the page like a human does. This makes it particularly strong for automating across diverse websites where you can’t control the markup — government forms, insurance portals, supplier sites.

Setup

Cloud (recommended — no local install needed):

  1. Get an API key from app.skyvern.com
  2. Configure your MCP client:
{
  "mcpServers": {
    "skyvern": {
      "command": "uvx",
      "args": ["skyvern", "run", "mcp"],
      "env": {
        "SKYVERN_API_KEY": "your-api-key",
        "SKYVERN_BASE_URL": "https://api.skyvern.com"
      }
    }
  }
}

Self-hosted (local mode):

Requires Python 3.11+:

pip install skyvern
skyvern quickstart
skyvern run server  # start local server

Then configure with SKYVERN_BASE_URL=http://localhost:8000.

Also available via one-click .mcpb bundle for Claude Desktop (macOS/Windows). Supports Claude Desktop, Claude Code, Cursor, Windsurf, and Codex.

What’s Good

Vision-based resilience. The computer vision approach genuinely solves the biggest pain point in browser automation — selectors that break when UIs change. For scraping sites you don’t control, this is transformative.

Natural language actions. skyvern_act("fill out the contact form with my info") is vastly simpler than writing selector chains. The LLM figures out what elements to interact with and in what order.

Credential management. Native integration with Bitwarden, 1Password, and Azure Key Vault, plus automatic 2FA/TOTP handling. Logging into sites with MFA is one of the hardest browser automation problems, and Skyvern handles it out of the box.

WebBench SOTA for write tasks. 64.4% overall accuracy on WebBench (5,750 tasks across 452 websites), and best-performing agent on WRITE tasks (form filling, login, downloads). 85.85% on the older WebVoyager benchmark.

Dual deployment model. Cloud for zero-setup production use with anti-bot protection and proxy networks. Self-hosted for privacy-sensitive use cases where browser data shouldn’t leave your network.

QA skill. Local mode includes a /qa feature that reads git diff, generates test cases, and runs automated browser tests against your dev server — practical for frontend developers.

What’s Not

Cost at scale. Cloud pricing is $0.10/page with credits-based tiers: Free (1,000 credits), Hobby ($29/month), Pro ($149/month), Enterprise (custom). Vision + LLM processing means each action is significantly more expensive than a raw Playwright command. High-volume scraping gets costly fast.

Latency. Computer vision processing adds latency compared to direct DOM interaction. Browser sessions take a few seconds to start, and each action requires an LLM inference step. Not suitable for speed-critical automation.

Cloud dependency for best experience. Anti-bot protection, proxy networks, and parallel execution are cloud-only features. Self-hosted mode requires you to bring your own LLM API keys and handle infrastructure.

Accuracy ceiling. 64.4% on WebBench means roughly 1 in 3 tasks fail. For complex multi-step workflows, compounding error rates can make end-to-end reliability challenging. Compare to Anthropic Sonnet 3.7 CUA (current SOTA) or Surfer 2 (97.1% on WebVoyager).

Python 3.11+ requirement. Local mode won’t run on Python 3.10 or earlier, which may exclude some environments.

Active development velocity declining. Release notes show v1.0.31 dated April 14, 2024 (sic — likely a metadata issue), but the 11-person team and YC backing suggest continued investment. Watch for commit frequency trends.

Security Considerations

Skyvern’s security model has both strengths and gaps:

Strengths:

  • Credential vault integration (Bitwarden, 1Password, Azure Key Vault) — credentials never stored in plaintext config
  • Auth state persistence avoids repeated logins
  • Cloud mode isolates browser sessions from your local machine
  • Security policy at security@skyvern.com for responsible disclosure

Concerns:

  • The skyvern_evaluate tool runs arbitrary JavaScript in the browser context — standard browser automation risk but worth noting
  • Local mode with CDP connection to your actual browser exposes your browsing session to the MCP client
  • Anti-bot and CAPTCHA solving features exist in a gray area — useful for legitimate automation, but could facilitate ToS violations
  • No Skyvern-specific CVEs found, but the broader browser automation MCP category has known attack surfaces (prompt injection via page content, session hijacking)
  • CodeBlock sandbox received a critical fix hardening against dangerous attribute access — suggests the attack surface was real

Recommendation: Use cloud mode for production and sensitive workflows. If self-hosting, run in an isolated environment with dedicated browser profiles.

How It Compares

vs Playwright MCP (31K stars, 4/5): Playwright is faster, free, and more deterministic. Skyvern is more resilient to UI changes and easier to write natural language automations for. Playwright for known, stable sites; Skyvern for diverse or changing sites.

vs Browserbase MCP (3.3K stars): Browserbase provides cloud browser infrastructure. Skyvern provides the AI automation layer. They’re complementary — Skyvern can run on Browserbase infrastructure.

vs Desktop Commander (5.9K stars, 3.5/5): Desktop Commander is local-only, terminal+filesystem focused. Skyvern is browser-focused with cloud option. Different tools for different jobs.

vs Browser-Use (78K stars, no review yet): Browser-Use has much larger community (3.8x stars) and higher benchmark accuracy (89.1% vs 64.4% on WebBench). However, Browser-Use is primarily a Python library, not a dedicated MCP server — Skyvern has a more polished MCP integration and managed cloud offering.

Who Should Use This

Skyvern is best suited for:

  • RPA/automation teams automating across diverse third-party websites (insurance forms, government portals, supplier systems)
  • Web scraping where sites change frequently and selector-based scrapers break
  • QA teams wanting AI-powered end-to-end testing without maintaining brittle test selectors
  • Developers who need credential-managed login automation with 2FA support

It’s overkill for automating your own app (use Playwright) or simple single-site scraping (use Firecrawl).

Bottom Line

Skyvern brings a genuinely novel approach to browser automation — using computer vision instead of selectors is the right long-term bet for cross-site automation. The 75+ tool MCP server is comprehensive, the credential management is best-in-class, and the dual cloud/self-hosted model gives deployment flexibility. The YC backing and $2.7M in funding provide some assurance of continued development.

The tradeoffs are real: higher cost per action than raw Playwright, latency from LLM inference, and a 64.4% accuracy ceiling that means you’ll need error handling and retries for complex workflows. But for the use cases where selector-based automation is fundamentally fragile — diverse sites, changing UIs, authenticated workflows — Skyvern is the strongest MCP option available.

Rating: 4/5 — Best-in-class vision-based browser automation with strong MCP integration and managed cloud offering. Docked for cost at scale, accuracy ceiling below competitors, and cloud dependency for full feature set.

Last updated: April 20, 2026