ghost-browser

# Ghost Browser — AI-Powered Browser Automation > The browser moves on its own. No human needed. **GitHub:** https://github.com/aiagentwithdhruv/ghost-browser **Path:** `Social-Media-Agent-1.0/browser-automation/` **Stack:** Python 3, Playwright, OpenAI API --- ## What It Does Automates any browser action with human-like behavior — LinkedIn posting, engagement, job applications, web scraping, stats tracking, and visual demos. --- ## Schema ### Inputs | Name | Type | Required | Description | |------|------|----------|-------------| | `LINKEDIN_LI_AT` | string | For LinkedIn | LinkedIn `li_at` cookie from DevTools | | `TWITTER_AUTH_TOKEN` | string | For Twitter | Twitter auth_token cookie | | `TWITTER_CT0` | string | For Twitter | Twitter ct0 cookie | | `YOUTUBE_API_KEY` | string | For YouTube | Google Cloud API key (free) | | `OPENAI_API_KEY` | string | For AI comments | GPT-4o-mini for comment/cover letter generation | ### Outputs | Name | Type | Description | |------|------|-------------| | LinkedIn post | Action | Published post with optional image | | Feed data | JSON | Posts from LinkedIn feed with author, content, URN | | Job listings | JSON | Jobs from LinkedIn/Indeed with title, company, description | | Scraped data | JSON/CSV | Data from any website (smart/full/selector modes) | | Stats | JSON + MD | Followers, views, connections across all platforms | | Screenshots | PNG | Any URL, full page or element | ### Credentials | Service | How to Get | |---------|-----------| | LinkedIn li_at | DevTools → Application → Cookies → linkedin.com | | Twitter auth_token + ct0 | DevTools → Application → Cookies → x.com | | YouTube API | Google Cloud Console → YouTube Data API v3 (FREE) | | OpenAI | platform.openai.com → API Keys | ### Cost | Component | Cost | |-----------|------| | Playwright/Chromium | Free | | YouTube API | Free (10K units/day) | | OpenAI (comments) | ~$0.01 per comment | | LinkedIn/Twitter | Free (cookie auth) | --- ## Scripts Reference | Script | What It Does | Example Command | |--------|-------------|----------------| | `orchestrator.py` | Multi-agent orchestrator (8 agents, 5 combos) | `python3 orchestrator.py --demo --visible` | | `ghost_mcp.py` | MCP server (10 tools for any AI agent) | `python3 ghost_mcp.py` (stdio transport) | | `multi_context.py` | Isolated browser session manager | Used by orchestrator | | `linkedin_engage.py` | LinkedIn CLI — post, feed, engage, apply, connect | `python3 linkedin_engage.py post --text "Hello" --image photo.jpg --visible` | | `universal_scraper.py` | Scrape any website with 6 presets | `python3 universal_scraper.py indeed --query "AI engineer" --output jobs.json` | | `stats_tracker.py` | Pull stats from all platforms | `python3 stats_tracker.py --notify` | | `demo_wow.py` | 6-stage visual demo (viral) | `python3 demo_wow.py` | | `monkeytype_flex.py` | Speed typing demo (412 WPM) | `python3 monkeytype_flex.py` | | `job_research.py` | Browse AI jobs on LinkedIn + Indeed | `python3 job_research.py` | | `screenshot_tool.py` | Screenshot any URL | `python3 screenshot_tool.py --url "https://example.com"` | --- ## Architecture (v2.0) ``` human_behavior.py ← Anti-detection layer (mouse drift, idle, warmup) ↓ base_browser.py ← Playwright wrapper (cookie auth, context manager) multi_context.py ← Isolated browser sessions (multiple contexts, 1 process) ↓ orchestrator.py ← 8 agents, 5 combos, CLI ├── agent_linkedin_feed / agent_linkedin_engage ├── agent_scrape_indeed / agent_scrape_gmaps ├── agent_twitter_trending (cookie auth: TWITTER_AUTH_TOKEN + TWITTER_CT0) ├── agent_monkeytype / agent_github_trending / agent_hacker_news └── run_orchestrator() → parallel execution in isolated contexts ↓ ghost_mcp.py ← MCP server (10 tools) for Claude/Cursor/Codex ↓ ├── linkedin_browser.py ← LinkedIn: post, like, comment, apply, connect ├── linkedin_scraper.py ← LinkedIn: followers, connections, analytics ├── twitter_scraper.py ← Twitter: followers, impressions ├── youtube_stats.py ← YouTube API (no browser) ├── universal_scraper.py ← Any website (6 presets) └── screenshot_tool.py ← Screenshots ↓ ├── linkedin_engage.py ← CLI for LinkedIn ├── stats_tracker.py ← Stats + Telegram notifications ├── demo_wow.py ← Viral demo └── job_research.py ← Job market research ``` ### Agents & Combos (v2.0) | Agent | Description | |-------|-------------| | `linkedin-feed` | Scroll feed, extract posts | | `linkedin-engage` | Like posts with human behavior | | `indeed` | Scrape Indeed job listings | | `gmaps` | Scrape Google Maps businesses | | `monkeytype` | Speed typing demo (~400 WPM) | | `github` | Browse GitHub trending repos | | `hackernews` | Read top Hacker News stories | | `twitter` | Scrape X.com trending/search (cookie auth) | | Combo | Agents | |-------|--------| | `demo` | monkeytype, github, hackernews | | `research` | linkedin-feed, indeed, github, twitter | | `lead-gen` | linkedin-feed, gmaps, indeed | | `social` | linkedin-feed, twitter, hackernews | | `full` | all 6 non-demo agents | ### MCP Server (10 tools) ```json { "mcpServers": { "ghost-browser": { "command": "python3", "args": ["/path/to/ghost_mcp.py"], "env": { "LINKEDIN_LI_AT": "...", "OPENAI_API_KEY": "sk-..." } } } } ``` Tools: linkedin_feed, linkedin_post, linkedin_engage, scrape_url, scrape_indeed, scrape_gmaps, scrape_twitter, typing_demo, screenshot, multi_run --- ## Human-Like Behavior Rules | Behavior | Implementation | |----------|---------------| | Typing | 40-55 WPM, char-by-char, 3-5% typo rate with backspace correction | | Delays | Gaussian distribution (not uniform) — clusters around midpoint | | Scrolling | Sinusoidal speed curve — slow start, fast middle, slow end | | Clicking | Random offset from center (not dead-center like bots) | | Reading | Wait time proportional to word count | | Breaks | 30-120 sec pause every 15-25 actions | | Viewport | Random from pool: 1440x900, 1920x1080, 1536x864, etc. | --- ## LinkedIn Post Flow ``` 1. Navigate to feed 2. Click "Start a post" 3. Find Quill editor (contenteditable div) 4. Short posts (<200 chars) → human_type with typos Long posts (>200 chars) → insert_text (paste) 5. Attach image → file input upload 6. Click "Next" (LinkedIn image editor step) 7. Click "Post" → confirmed ``` --- ## Composable With | Skill | How | |-------|-----| | `scrape-leads` | Scrape → enrich with ghost-browser LinkedIn data | | `instantly-campaigns` | Scrape leads → cold email | | `gmaps-leads` | Google Maps preset in universal_scraper | | `cross-niche-outliers` | Research content → auto-post to LinkedIn | | `upwork-apply` | Job research data feeds Upwork proposals | | `send-telegram` | Stats tracker sends notifications | --- ## Revenue Use Cases | Use Case | Revenue | |----------|---------| | LinkedIn automation service | $500-2000/mo per client | | Web scraping gigs (Fiverr/Upwork) | $200-1000 per project | | LinkedIn ghostwriting + auto-post | $500-1500/mo per client | | AI job application bot | $50-200 per person | | Lead generation (scrape → email) | $500-2000/mo per client | --- ## Quick Start ```bash cd Social-Media-Agent-1.0/browser-automation pip install -r requirements.txt playwright install chromium cp .env.example .env # Add your cookies # Run the viral demo python3 demo_wow.py # Post to LinkedIn python3 linkedin_engage.py post --text "Hello world" --visible # Scrape Indeed jobs python3 universal_scraper.py indeed --query "AI engineer" --output jobs.json ```