Browser Harness: AI's Full Browser Control via CDP
Browser Harness repo uses Chrome DevTools Protocol for precise mouse/keyboard simulation, self-updates its helpers.py for new tasks, and pre-builds skills for sites like TikTok/Zillow—founders bet a Mac Mini on any failure.
Direct CDP Access Simulates Human Browser Interaction
Browser Harness connects LLMs directly to Chrome's DevTools Protocol (CDP) for granular control over tabs, uploads, downloads, drag-and-drop, iframes, and more, mimicking mouse and keyboard inputs without subscriptions. Run locally by cloning the public GitHub repo, allowing remote debugging in Chrome, or use free cloud option with 3 concurrent browsers (no credit card needed). This bypasses anti-bot measures like Cloudflare by executing human-like actions, enabling reliable web scraping of obfuscated data such as emails hidden behind 'reveal' buttons on directories.
Setup launches Chrome automatically; add your API key (e.g., OpenRouter, Claude) for LLM integration. Interaction skills list covers all common browser actions, letting agents handle complex flows like generating videos in tools like C-dance 2, uploading to TikTok Studio, scheduling posts, and analyzing view-based hooks.
Self-Annealing Helpers.py Enables Permanent Skill Acquisition
The helpers.py file acts as a dynamic toolkit: when an agent encounters a new task (e.g., file upload), it writes the required Python code directly into helpers.py, making the capability persistent across sessions. This self-annealing process shares domain-specific skills between agents, eliminating per-task reprogramming.
Pre-built Markdown files provide ready harnesses for platforms including TikTok, Facebook, Zillow (real estate listings), eBay, Etsy, Craigslist, and Steam. Agents modify their own harnesses for impossible tasks, as demonstrated by founder screenshots: Claude drawing hearts in Excalidraw or full TikTok video pipelines.
AI as Manager for Deterministic Scraping Pipelines
Combine Browser Harness with Python scripts for production reliability: AI prototypes a standard operating procedure (SOP), then converts it to a deterministic script that runs autonomously, collecting data (e.g., revealed emails) into CSV. Wake AI only on errors like blank pages or lags via browser harness integration; AI fixes issues or escalates via SMS/Telegram.
This shifts AI from micro-managing to oversight, ethical for public data (no server strain, simulates manual collection). Scales to social media automation on hard-to-API platforms like Instagram/LinkedIn/Facebook, running free locally with Claude or Py LLMs. Founders' confidence: first failed task wins a Mac Mini.