Show HN: Vibium – Browser automation for AI and humans, by Selenium's creator

Original Article ↗ Hacker News Discussion ↗

Positioning vs Existing Tools (Playwright, Selenium, others)

  • Multiple commenters ask what differentiates Vibium from Playwright, since both are “batteries included” and AI-friendly.
  • Author frames Vibium as built “AI-first”: not just an action API but a full sense–think–act loop for agents.
    • V1 = “clicker” (act: click/type/navigate/screenshot).
    • V2 plans = “retina” (sense, turning interactions/tests/prod usage into durable signals) and “cortex” (think, building reusable workflow models so agents don’t reason from raw HTML every time).
  • Vibium relies heavily on the WebDriver BiDi standard; goal is to embrace what Playwright does well and then extend beyond it.
  • Selenium is described as still massively entrenched, especially outside startups; Vibium is pitched as a future bridge for that installed base.

Why a New Project, Not “Selenium + AI”

  • Technical: Selenium and Playwright weren’t designed with AI workflows in mind from day one.
  • Branding: Selenium has accumulated “baggage”; many newer developers assume “Playwright good, Selenium bad,” making a fresh name more viable.

Current Capabilities and Roadmap

  • Today:
    • Go binary (“clicker”), MCP server, JS/TS API.
    • Can be used without any LLM or via MCP (e.g., in Claude Code).
  • Missing but requested features:
    • JS eval and DOM access via MCP, network monitoring/modification, richer browser state exposure, better navigation from screenshots, handling logins/cookies beyond basic test setups.
  • Planned:
    • Python and Go clients, Agent Skills support, local model support (e.g., via llama.cpp), tighter browser controls (URL allowlists, network rules), stronger security models, and possibly hosted “vibium.cloud”.

Security, Governance, and Sandboxing

  • Strong concern about the “blast radius” of agent-driven browsers.
  • Suggested mitigations: VMs, containers with strict firewalls/URL allowlists, MCP hooks to gate navigation, defense-in-depth concepts (logging, policies, protector agents, “panic modes”).

Use Cases, Adoption, and Limitations

  • Early users employ Vibium with Claude to drive local web apps and summarize content.
  • Some focus on “just clicking” for tests and workflows; others stress they mainly need JS injection and network inspection and see current v1 as too limited.
  • Captchas and fragile UI flows remain problematic; some hope Vibium will handle these better, others are skeptical.
  • Overall sentiment mixes enthusiasm (especially from long-time Selenium users) with calls for clearer, concrete differentiation and more mature features in v2+.