human-mcp

# Human MCP **AI Agent fully automated image search and social media posting tool.** From Google Trends / Weibo hot search to grab keywords → Playwright headless image download → Gemini generate captions → automatically post to Facebook + Threads + Instagram. --- ## Core Architecture: Semantic Clicking ### What Problem Does It Solve? Traditional RPA scripts use coordinate clicking `click(500, 300)`, which breaks when the button changes position/color/style. Human MCP's approach: **from "coordinate-oriented" to "semantic visual-oriented"**. ### Three-Layer Self-Healing Mechanism ``` ┌─────────────────────────────────────────────────────┐ │ Layer 1: DOM Positioning (getByRole / getByText) │ │ → UI changes? As long as role + label are still there │ │ ↓ can't find │ │ Layer 2: JS dispatchEvent (React compatibility fallback) │ │ ↓ no response │ │ Layer 3: Result Verification + Retry (Self-Correction Loop) │ │ → verify_after keyword, incorrect automatically retries up to 3 times │ └─────────────────────────────────────────────────────┘ ``` Each click step has verification + retry, effective even after IG/Threads UI updates. ```python # Core: SemanticClicker ok = await clicker.click( label="Post", role="button", parent="dialog", verify_after="Posted", # verify keyword after clicking max_retries=3, # automatically retry on error ) ``` ### Why Not Use CDP Vision? CDP (Remote Debugging Protocol) relies on DOM to find coordinates, which breaks when UI changes. Semantic clicking uses `getByRole` + `getByText`, which are DOM-level semantic labels, much more stable than coordinates. Vision is only reserved as a last resort fallback. --- ## Independent Chromium Profile Each platform uses an **independent Chromium Profile**, which does not affect the user's normal browser: | Script | Profile Directory | Purpose | |------|------------------|------| | `post_threads.py` | `/tmp/threads-chromium-profile/` | Threads image and text posting | | `post_ig_human.py` | `/tmp/ig-chromium-profile/` | Instagram image and text posting | | `post_facebook.py` | `/tmp/fb-chromium-profile/` | Facebook image and text posting | First-time login required, then automatically remembers session. --- ## Quick Start ```bash # First-time (only once per platform) python3 post_threads.py "Test Post" "/path/to/image.jpg" python3 post_ig_human.py "Test Post" "/path/to/image.jpg" python3 post_facebook.py "Test Post" "/path/to/image.jpg" # Full workflow python3 social_workflow.py 1 # Weibo hot search → FB + Threads + IG python3 social_workflow.py 2 # Google Trends HK python3 social_workflow.py 3 # Google Trends US ``` ## API (Image Download) ```bash uv run python server.py # start server curl "http://localhost:8080/scrape?query=jimmy+kimmel&engine=bing&max_images=3" curl "http://localhost:8080/download?url=https://..." ``` ## Installation ```bash cd ~/human-mcp npm install pip install fastapi uvicorn ```