Show HN: Browser Harness – Gives LLM freedom to complete any browser task

· ai ai-agents coding · Source ↗

TLDR

  • Self-healing Python harness (~592 lines) connecting LLMs directly to Chrome via CDP, where the agent writes missing helpers mid-task.

Key Takeaways

  • Built on raw CDP with one WebSocket to Chrome; no framework, no recipes, no abstraction layer between agent and browser.
  • Self-healing means the agent edits helpers.py itself when a needed function is missing – e.g. upload_file() gets written live during a run.
  • Core files are thin by design: run.py (~36 lines), helpers.py (~195 lines), daemon/CDP bridge (~361 lines).
  • Domain skills under domain-skills/ (GitHub, LinkedIn, Amazon) are agent-generated, not hand-authored – the harness files them automatically after solving a non-obvious flow.
  • Free remote browser tier at cloud.browser-use.com offers 3 concurrent browsers with proxies and captcha solving, no card required.

Hacker News Comment Review

  • Sparse discussion with only two live comments; no deep technical debate on CDP internals, self-healing reliability, or comparison to Playwright/Puppeteer-based stacks.
  • One commenter confirmed similar results using Vercel’s agent-browser and Playwright for scraping and automation, suggesting the approach is replicable with other drivers.
  • “Open washing” skepticism surfaced – the project’s openness or licensing framing drew mild critique, though no specific closed component was identified.

Notable Comments

  • @syl5x: Achieved comparable results with Vercel agent-browser and Playwright; calls it good for scraping and browser automation.
  • @Amekedl: Flags possible “open washing” while calling the project cool.

Original | Discuss on HN