moates-mcp

An MCP server that answers questions about Marcus Oates, a Sydney-based senior software engineer (voice AI & backend).

It's a proof of concept: any MCP client (Claude Desktop, Claude Code, an Agent SDK app) connects to the server and uses its tools to answer questions about Marcus. The client brings the LLM; this server just serves accurate, structured data (mirrored from https://moates.com.au) so the model doesn't have to guess.

Tools

Tool	Returns
get_profile	Name, title, location, summary, contact links
list_projects(status?)	Projects, optionally filtered by prod / test / poc
get_project(key)	Full detail for one project
get_experience	Work history with detailed highlights
get_skills	Tech stack, grouped
get_interests	Life outside work
get_resume	The whole profile as one markdown document
search(query)	Keyword search across experience, projects and skills

Also exposes a resume://marcus resource and an ask_about_marcus prompt.

Run locally

uv venv --python 3.12
uv pip install -e ".[dev]"

# stdio (quick local test, e.g. from Claude Desktop config)
uv run moates-mcp

# Streamable HTTP on 127.0.0.1:8000/mcp
MCP_HOST=0.0.0.0 uv run python -m moates_mcp --http

Run the tests with uv run pytest.

Connect a client

Claude Code (HTTP):

claude mcp add --transport http marcus https://mcp.moates.com.au/mcp

Claude Desktop (stdio) — add to claude_desktop_config.json:

{
  "mcpServers": {
    "marcus": { "command": "uv", "args": ["run", "moates-mcp"], "cwd": "/path/to/moates_mcp" }
  }
}

Deploy to a DigitalOcean droplet

The server speaks MCP's Streamable HTTP transport, so it deploys like any ASGI web app. See deploy/ for a Dockerfile, an nginx reverse-proxy config, and a systemd unit. In short:

1. Point a subdomain (mcp.moates.com.au) at the droplet.
2. Run the container (or the systemd service) listening on 127.0.0.1:8000.
3. Front it with nginx + a Let's Encrypt cert on 443, proxying to /mcp.

Clients then connect to https://mcp.moates.com.au/mcp.