Show HN: Open source alternative to Perplexity Comet

Original Article ↗ Hacker News Discussion ↗

Architecture and Form Factor

  • Built as a Chromium fork / “wrapper” with C++ patches rather than just an Electron/Playwright-style controller.
  • Key change is enriched DOM and accessibility trees exposed at the C++ level for fast agent actions (click, input, element lookup), claimed to be 20–40x faster than injected JS.
  • Team plans to expose these low-level APIs to developers and potentially implement MCP so the browser can act as a sub-agent for external orchestrators (e.g., automated testing).
  • Several users would prefer a browser extension; maintainers argue Chromium’s extension APIs don’t expose what they need (especially the accessibility tree) and are intentionally constrained.

Security, Maintenance, and Engine Choice

  • Strong concern that a tiny team cannot realistically track Chromium’s rapid, security-critical release cycle and associated CVEs.
  • Critics emphasize the RCE risk of lagging even a few weeks behind upstream; they ask whether every release branch will be rebuilt and shipped.
  • Maintainers say they currently build on the same Chromium release as Chrome, have an auto-updater, and plan to scale the team; Brave is cited as proof a fork can be viable.
  • Some question using Chromium for a “privacy-first” product instead of Firefox; maintainers respond that non-Chromium engines require huge effort to achieve site compatibility and extension support.

Privacy, Models, and Local vs Cloud

  • Pitch is privacy-first: agent runs locally, users can bring their own API keys, and local LLMs via Ollama are supported.
  • Default today is Gemini via a liteLLM proxy (key hidden server-side); a smaller, fine-tuned local model is promised as the future default.
  • Plans include a built-in adblocker plus LLM-based detection of more ad formats; skeptics question using LLMs per-request.

Use Cases, UX, and Current Limitations

  • Vision: automate repetitive workflows (e.g., ecommerce repricing, complex trip planning).
  • Some users say existing tools (Playwright scripts, traditional sites like Orbitz) often solve these better today.
  • Demo (buying toothpaste) is criticized as underspecified and not clearly time-saving; calls for more challenging, realistic demos and better failure recovery.
  • One tester reports a comment-summarization task taking >20 minutes, requiring manual guidance, with many small scroll actions—currently too slow and fragile for their use.
  • Requests for clearer UI affordances (visible cursor movement, typed keys) to make agent behavior understandable.

Business Model and Ecosystem

  • Revenue plan: free consumer browser, paid enterprise edition; AGPL licensing is seen as consistent with that.
  • Linux support is repeatedly requested; maintainers say it’s imminent.
  • Some see this as the “real” open-source version of Perplexity Comet and welcome transparent, local-first agents; others dismiss it as minimal patches over a huge upstream codebase and a “solution looking for a problem.”