Skyvern
Open-source browser-automation agent that uses vision + LLM reasoning to handle dynamic sites scrapers can't.
Delv Safety Grade: B
Score 72/100 · assessed 2026-04-18
Skyvern is an open-source browser automation agent from a venture-backed startup (Skyvern Inc.) that uses vision models and LLMs to navigate dynamic websites. The maintainer is a small but funded team with active development since 2024. Transparency is strong: full source on GitHub, clear documentation, active issue tracker. Supply chain is reasonable via PyPI packages and Docker images with versioned releases. The major safety concern is the permission surface: Skyvern requires full browser control, can execute arbitrary workflows described in natural language, and typically needs access to credentials for the sites it automates. This is inherent to its purpose but creates significant risk if misconfigured or if malicious workflows are fed to it. No known security incidents. Best suited for controlled environments with careful workflow review.
Green flags
- Fully open source with active GitHub repo and clear commit history
- Well-documented architecture and deployment guides
- Distributed via standard PyPI and Docker Hub with versioned releases
- Active community and responsive issue tracker
- No known security incidents or malicious use cases reported
Red flags
- Full browser control with arbitrary site navigation and form submission
- Requires credentials for target sites, creating credential exposure risk
- LLM-driven actions can be unpredictable on adversarial or malformed sites
- Self-hosted deployment means security responsibility falls on operator
- Small team with limited bus factor for security maintenance
Permissions requested
Pricing
Platforms
Review
Pay for Skyvern if you automate forms on sites that change often or if compliance demands self-hosting. Skip it if you need sub-second scraping or if your targets are static enough for Puppeteer.
Good at
- Survives DOM changes that break traditional scrapers
- Self-hosted option for sensitive data workflows
- LLM reasoning handles validation errors and retries intelligently
- Vision-based element detection works across layout shifts
- Open-source core with transparent pricing
Watch out
- 3-5 second latency per action makes high-speed scraping impractical
- Vision model occasionally misidentifies visually similar buttons
- Aggressive bot detection can still block the underlying browser session
- Requires more compute than traditional scrapers
- Learning curve steeper than visual workflow builders like Axiom
Use cases
- Scraping sites that fight scrapers
- Repeatable form-filling at scale
- Insurance and finance form automation
- Self-hosted browser agent