Wobbo Bobo Ops & Creative Sandbox — MCP Server

A Model Context Protocol server that complements (not replaces) the n8n Stage 1+2 automation. The n8n workflow owns the always-on, Telegram-triggered pipeline; this server gives a conversational window into the same Postgres state, plus a sandbox for testing the Direction/Hook/Script agents without a full Telegram round-trip.

What it adds

• A missing piece the n8n pipeline has no home for: logging TikTok view counts back against a project, and correlating performance by creative direction / hook style / video-gen model — the actual data your MVP-exit decision (100k views) depends on.

• Conversational management of the off-limits rules and hook-styles lists (same Postgres tables the n8n workflow reads/writes), instead of the Supabase SQL editor.

• A sandbox to call the Direction/Hook/Script agents directly and iterate on prompts, without waiting on a Telegram approval loop.

Related MCP server: General-Purpose MCP Database Server

Setup

cd mcp-server
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# edit .env: DATABASE_URL (same Supabase session-pooler string used in n8n),
# ANTHROPIC_API_KEY (same key used in the n8n Anthropic credential)

Then run the schema addition once (Supabase SQL Editor):

cat schema.sql

Run it standalone (smoke test)

source .venv/bin/activate
python3 server.py

It should sit waiting on stdio — that's normal, it's meant to be driven by an MCP client, not run interactively. Ctrl+C to stop.

Register with Claude Code

claude mcp add wobbobobo -- python3 /Users/eugeneshemchuk/DEV/n8n/mcp-server/server.py

(Run claude mcp add --help if the flags above don't match your installed version — the add subcommand's exact syntax has changed across releases.)

Register with Claude Desktop

Add to your claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "wobbobobo": {
      "command": "python3",
      "args": ["/Users/eugeneshemchuk/DEV/n8n/mcp-server/server.py"]
    }
  }
}

Restart Claude Desktop after editing.

What's exposed

Resources (readable data):

• wobbobobo://off-limits-rules — active off-limits rules

• wobbobobo://hook-styles — active rotating hook styles

• wobbobobo://project/{project_id} — a single project's full record

Tools:

• list_projects(status?, limit?) — browse projects by status

• get_project(project_id) — full detail on one project

• log_video_performance(project_id, views) — log a TikTok view count

• correlate_performance(group_by) — best-views breakdown by direction / hook_style / model_used

• add_off_limits_rule(rule_text) / deactivate_off_limits_rule(id)

• add_hook_style(style_name, style_description)

• generate_direction(idea, verdict?, reasoning?) — sandbox call to the Direction Agent

• generate_hooks(idea, direction_tone, visual_style, why_it_could_hit) — sandbox call to the Hook Agent

• generate_script(idea, direction_tone, visual_style, why_it_could_hit, hooks) — sandbox call to the Script Agent

• generate_script_edit(...) — sandbox call to the Script Agent's edit pass

Known tradeoff

The Direction/Hook/Script prompts here are copied from the n8n workflow's HTTP nodes, not shared from a single source. If you tune a prompt in one place, update the other by hand — there's no shared package linking them.