Bumblewick Support-Desk MCP Server
Enables AI assistants to manage customer support for a fictional candle shop, including order lookup, customer file access, and safe refund processing with server-side safety rules.
README
Bumblewick Support-Desk MCP Server 🕯️
The example project for the "Give the Robot Hands" series — a small, safe, well-tested Model Context Protocol server for a fictional online candle shop, Bumblewick & Co.
It gives an AI assistant carefully-scoped "hands": it can look up orders, read a customer's file, search the help centre, and issue refunds — with every dangerous action fenced in by server-side rules. No real people, no real money.
Everything runs locally against a seeded SQLite database. Nothing leaves your machine.
What it exposes
Tools
find_orders(query)— look up orders by order id, customer id, or email (read-only)get_order(order_id)— one order, its refund history, and the refundable balance (read-only)search_help(query)— search the help centre (read-only)create_refund(order_id, amount_cents, reason, confirm=false, idempotency_key=None)— issue a refund (guarded write)
Resources
bumblewick://policy/refunds— the refund policy, as contextbumblewick://customer/{customer_ref}— a customer file (profile + orders)
Prompts
draft_apology(order_id, tone="warm")— a reusable support-reply template
The safety rules (a.k.a. "don't give the robot a chainsaw")
create_refund refuses to misbehave, server-side:
- Only shipped/delivered orders are refundable.
- Never more than the remaining refundable balance (no over-refunds).
- Amounts over the auto-approve limit return
needs_confirmationand do not execute until you re-call withconfirm=true(human-in-the-loop). - A per-customer rate limit caps refunds within a rolling window.
- An idempotency key makes retries safe — the same key returns the original refund, and reusing a key with different parameters is rejected (never a silent wrong-refund).
- The check-and-write runs in one atomic transaction (
BEGIN IMMEDIATE), so two concurrent refunds can't both slip past the balance check. - A non-empty reason is required, and every decision is written to an audit log on stderr (stdout is reserved for the JSON-RPC stream).
Quick start
# 1. Install (any of: pip, uv, poetry)
pip install -e ".[dev]"
# 2. Run the tests
pytest
# 3. Explore it in the MCP Inspector (opens a GUI to poke every tool)
mcp dev src/bumblewick_support/server.py
# 4. Or run it over stdio (for a client like Claude Desktop)
python -m bumblewick_support.server
First run seeds bumblewick.db with sample customers, candle orders, and help
articles. Delete the file to reset.
Connect it to Claude Desktop
Add this to your claude_desktop_config.json:
{
"mcpServers": {
"bumblewick": {
"command": "python",
"args": ["-m", "bumblewick_support.server"]
}
}
}
Then ask: "Find Ivy's orders and refund the Vanilla Doom candle." Watch it use
find_orders, get_order, and create_refund — and watch it get politely
stopped when it tries to refund something it shouldn't.
Configuration
All settings are environment variables (prefix BUMBLEWICK_); see .env.example.
| Variable | Default | Meaning |
|---|---|---|
BUMBLEWICK_DB_PATH |
bumblewick.db |
SQLite file path |
BUMBLEWICK_AUTO_REFUND_LIMIT_CENTS |
2000 |
auto-approve ceiling; above needs confirm=true |
BUMBLEWICK_REFUND_RATE_LIMIT |
3 |
max refunds per customer per window |
BUMBLEWICK_REFUND_RATE_WINDOW_HOURS |
24 |
rate-limit window |
Project layout
src/bumblewick_support/
server.py # thin MCP layer: tools, resources, prompts
services.py # SupportDesk — all business logic (testable, no MCP)
safety.py # pure refund-policy engine
db.py # SQLite data access
models.py # pydantic domain models
seed.py # deterministic sample data
config.py # env-driven settings
tests/ # pytest: safety, services, and server smoke tests
The golden rule the series teaches: keep the business logic out of the MCP
layer. server.py only wires things up; everything worth testing lives in
services.py and safety.py, so it's testable without a client.
License
MIT.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
Neon Database
MCP server for interacting with Neon Management API and databases
E2B
Using MCP to run code via e2b.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.