Rheintal Pumpen

A public MCP server that simulates a fictional German B2B mid-market company – with role-based access control, audit logging, and a clean architecture you can study.

Connectable to Claude Desktop, a local LibreChat instance, or your own applications.

License: MIT Python 3.11+

What is this?

A demonstration of how standardised AI tool integration can look in a mid-market company. Behind the MCP server sits a fictional firm, Rheintal Pumpen GmbH in Mainz, Germany – an industrial-pump maker with ~180 employees, 60 customers, 250 orders, plus complaints and service tickets.

Bring-your-own-key, self-hostable, MIT licensed.

Why is the company name German? Rheintal Pumpen GmbH is the story of the demo (a German Mittelstand company). Everything around it – code, tool descriptions, status enums, documentation – is in English.

Two ways to try it

Full local demo: LibreChat (15 minutes)

git clone https://github.com/dataAiOliver/rheintal-pumpen-mcp.git
cd rheintal-pumpen-mcp
make up

make up does everything in one step: creates .env from .env.example, generates the SQLite database if it doesn't exist, and starts LibreChat + MCP + MongoDB + Meilisearch via Docker Compose.

The first run prints a hint telling you where to put your OpenAI or Anthropic key in .env. After that, run make up again and you're set.

Open http://localhost:3080 → register → start chatting.

You also get the built-in MCP Inspector at http://localhost:8765/: tool list, JSON schemas, sample calls. Machine-readable as /tools.json.

Live Activity at http://localhost:8765/activity: every tool call pops up live as a card with role, arguments and result. Nice demo punch line – stakeholders see in real time that the LLM is operating audit-ably.

→ docs/quickstart-local.md

Your own server (30 minutes)

make deploy SERVER=user@host

Prerequisites: a Linux server with Docker, your own domain, ports 80/443 open. Traefik takes care of TLS via Let's Encrypt.

→ docs/quickstart-own-server.md

Verifying the customer setup

After make up you can spot-check the stack:

make smoke-test        # HTTP probes: /healthz, /tools.json, LibreChat root
make up-verify         # 'make up' followed by smoke-test

To prove that the entire customer journey works on a clean Linux box – without trusting your own pre-installed Docker, make, etc. – run the DinD harness:

make test-fresh-ubuntu

It spawns a privileged ubuntu:24.04 container, runs scripts/install-prereqs.sh to install Docker + make + git + curl + jq + python3 from apt, starts an inner Docker daemon, performs make up, and runs the smoke test. Takes 4–7 minutes (the LibreChat image pull dominates) and is fully self-cleaning. See scripts/.

Try these prompts

1. "Which customers have open invoices over 10,000 EUR?"
2. "Show me all complaints from the last quarter."
3. "Which submersible pumps with a stainless steel housing are in stock?"
4. "What was the largest order in Q1, and how was it shipped?"
5. "List the top 5 customers by revenue."
6. "Which service tickets have been open for more than 7 days?"
7. "Show me the order history of customer RH-K-00012."

→ docs/demo-scenarios.md

Architecture

┌──────────────┐  HTTP/SSE   ┌──────────────────────┐
│ Claude       │◄──────────► │ MCP server           │
│ Desktop      │             │  (Python, mcp SDK)   │
│ LibreChat    │             │                      │
│ Custom app   │             │  Tools  Auth  Audit  │
└──────────────┘             │           │          │
                             │           ▼          │
                             │      SQLite (mock)   │
                             └──────────────────────┘

→ docs/architecture.md

What is demonstrated

• MCP tool definitions with Pydantic schemas
• Role-based access control (sales, service, accounting, admin)
• Audit logging suitable for AI-Act / GDPR-style traceability
• Bring-your-own-key architecture
• Self-hosting with Docker Compose
• Clean separation of transport / logic / data
• Documented migration path SQLite → Postgres

Available MCP tools

Tool	Roles	Description
lookup_customer	sales+	Look up a customer by ID or company name
list_customer_orders	sales+	Orders of one customer
list_top_customers_by_revenue	sales+	Top-N customers, aggregated in a single call
get_order_details	sales+	Order incl. line items
get_invoice	sales+	One invoice
list_open_invoices	accounting, admin	Open/overdue invoices incl. customer name
get_delivery_status	sales+	Delivery and tracking status
search_products	all	Product search
get_product_details	all	Product master data
list_complaints	service, admin	Filter complaints
get_audit_log	admin	Inspect the audit log
create_service_ticket	service, admin	write, requires confirm: true
flag_invoice_for_review	accounting, admin	write, requires confirm: true

Architecture Decision Records

• 001 – Why SQLite
• 002 – Why MCP instead of REST
• 003 – Stdio and HTTP side-by-side
• 004 – Bring your own key
• 005 – No TLS on localhost

Licence

MIT. Use it. Fork it. Build on top.

About the author

Oliver Köhn – AI architect for the German Mittelstand.

LinkedIn · Contact