mark-coach-mcp

Local MCP server that turns Mark Builds Brands' YouTube knowledge into an AI coaching assistant for ecommerce and Facebook Ads.

Ask Claude to analyze your Ads Manager screenshots, debug campaigns, or get strategic advice — and it'll respond using Mark's actual frameworks, vocabulary, and mental models from his 100+ videos.

How it works

1. Downloads transcripts from Mark's YouTube channel
2. Indexes them locally into a vector database (ChromaDB)
3. Exposes a search_mark_knowledge tool via MCP
4. A Claude skill activates the persona and queries the knowledge base

All data stays on your machine. No API keys required.

Install (one line)

curl -fsSL https://raw.githubusercontent.com/BlueNacho/mark-coach-mcp/main/install.sh | bash

That's it. The installer will:

1. Clone the repo to ~/mark-coach-mcp
2. Install uv (Python package manager) if needed
3. Install yt-dlp if needed
4. Install Python dependencies
5. Download Mark's transcripts (~100 videos, a few minutes)
6. Index everything into a local vector DB
7. Auto-detect Claude Code and/or Claude Desktop on your machine and configure both
8. Install the mark-coach skill globally for Claude Code

After it finishes, restart Claude Desktop / Claude Code and start using /mark-coach.

Want it somewhere else? INSTALL_DIR=~/projects/mark-coach-mcp curl -fsSL ... | bash

Manual install

git clone https://github.com/BlueNacho/mark-coach-mcp ~/mark-coach-mcp
cd ~/mark-coach-mcp
./setup.sh

Using it

Claude Code

The skill is installed globally. Just open Claude Code and type:

/mark-coach

Or describe an ecommerce / Facebook Ads question and the skill activates automatically.

Claude Desktop

After restart, the search_mark_knowledge tool is available. To get the persona/voice, create a Project in Claude.ai (or Cowork Space) and paste the contents of skills/mark-coach.md into the project's Custom Instructions.

Adding new videos

When Mark publishes new content:

cd ~/mark-coach-mcp
./setup.sh

The setup script is idempotent — it re-downloads only new videos and re-indexes only the new ones (existing indexed videos are skipped via data/processed.txt).

Using your own transcripts

If you have your own .vtt files (from any creator), drop them into transcripts/ and run:

./setup.sh

The indexer accepts any YouTube .vtt file — it strips timestamps, dedupes, and chunks the text.

Project structure

mark-coach-mcp/
  install.sh         ← one-line bootstrap (clones repo, runs setup.sh)
  setup.sh           ← installs deps, downloads/indexes, wires Claude
  pyproject.toml     ← Python dependencies (managed by uv)
  src/
    indexer.py       ← converts .vtt transcripts → ChromaDB
    server.py        ← MCP server with search_mark_knowledge tool
  skills/
    mark-coach.md    ← Claude skill / persona definition
  transcripts/       ← your .vtt files (gitignored)
  data/              ← ChromaDB vector store (gitignored)

Requirements

• macOS or Linux
• git (preinstalled on macOS via Xcode CLT, on Linux via your package manager)
• Claude Code and/or Claude Desktop

Everything else (uv, yt-dlp, Python deps) is installed automatically.

Troubleshooting

MCP says "Failed to connect" after a reboot

The MCP needs the absolute path to uv. The setup script handles this automatically; if you registered manually, make sure your claude mcp add command uses the full path (e.g. /Users/you/.local/bin/uv, not just uv).

Re-run setup

The setup script is safe to run any number of times. It detects what's already in place and skips it.

cd ~/mark-coach-mcp && ./setup.sh

Uninstall

claude mcp remove -s user mark-coach 2>/dev/null
rm -rf ~/.claude/skills/mark-coach
rm -rf ~/mark-coach-mcp

For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or ~/.config/Claude/claude_desktop_config.json (Linux) and remove the mark-coach entry from mcpServers.