Google Flow Editorial Poster Scraper & Automation

A Playwright-based automation pipeline for generating sports editorial graphic posters using Google Flow.

Project Architecture

newlab/
├── .gitignore
├── README.md
├── requirements.txt
├── main.py                   # Primary entry point
├── flow_scraper.py           # Legacy entry point wrapper
└── app/
    ├── __init__.py
    ├── config.py             # Configs, selectors, and directory paths
    ├── prompts.py            # FLOW_PROMPT text templates
    ├── utils/
    │   ├── __init__.py
    │   ├── clipboard.py      # Windows clipboard image copy helper
    │   └── image.py          # Directory image scanning logic
    └── automation/
        ├── __init__.py
        └── flow.py           # Playwright UI steps & automation pipeline

Related MCP server: playwright-fixer-mcp

Prerequisites

• Python 3.10+

• Google Chrome browser

• Pre-configured chrome_profile/ directory logged into Google

Setup & Execution with UV

This project is configured to use uv for fast, isolated Python package and environment management.

1. Initialize / Install Dependencies

uv sync
uv run python -m playwright install chromium

2. Run FastMCP Server

To run the FastMCP server via STDIO (for Claude Desktop, Cursor, or MCP clients):

uv run python -m app.mcp.server

or using FastMCP CLI:

uv run fastmcp run app/mcp/server.py

3. Run MCP Inspector Mode

To launch the MCP Server with interactive web-based Inspector:

uv run fastmcp dev inspector app/mcp/server.py

or via npx:

npx @modelcontextprotocol/inspector uv run python -m app.mcp.server

4. Connect with Claude Desktop

Add the following block to your %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "google-flow": {
      "command": "C:/Users/sohel/AppData/Roaming/Python/Python312/Scripts/uv.exe",
      "args": [
        "--directory",
        "c:/MCP Servers/flow-project",
        "run",
        "python",
        "-m",
        "app.mcp.server"
      ]
    }
  }
}

Usage & Workflow (Session Isolation)

The MCP server is fully session-aware, allowing multiple concurrent users or conversations to generate and edit projects in isolation.

• Strict Session Requirement: Every tool call (tool_create_project, tool_input_images, tool_generate_poster, tool_edit_poster, tool_poster_ratio_editor) requires an active MCP session context. This is handled automatically for normal MCP client connections. If session context is missing, the tool will raise a clear error.

• Reference Images: If you want to use reference images, call tool_input_images FIRST. It securely uploads images to a session-isolated folder on the server. tool_generate_poster will then automatically pick them up; there is no need to pass images_b64 directly to it anymore (though it is still accepted as an optional override).

• Per-Session Project State: Project state (project_url, image_edit_page_url, etc.) is tracked per-session in db/projects.json. Each tool's "use the latest project" fallback (when explicit URLs aren't passed) strictly targets the CALLING SESSION'S own projects. Concurrent users can safely use the server simultaneously without their projects, edits, or uploaded images interfering.

• Queueing Behavior: Image generation runs one at a time (not in parallel) due to the shared Chromium browser profile. Concurrent requests are safely queued, not dropped.

Direct Script Execution

Execute standard automation pipeline via main.py:

uv run python main.py