Gold Dust Lab · Infrastructure

Signal Chain

How a question travels from you to the shop’s data — and the one place the local qwen model does the thinking instead of Claude.


The chain

Every request runs left to right — and only after a Cisco Duo sign-in. Just one branch drops down to the local model.

Gold Dust Lab MCP signal chain You ask Claude, the MCP client in the cloud. To reach the MCP server Claude first signs in through Cisco Duo (SSO plus MFA) and gets a JWT, which it sends as a bearer token over HTTPS through nginx on the AWS box. The MCP server verifies the Duo JWT and exposes twelve tools; only the ask_gold_dust_lab tool branches down to the local Ollama qwen model. Cisco Duo SSO · MFA — issues JWT 1. sign in + MFA → JWT You Claude MCP client · brain cloud · Anthropic 2. Bearer JWT HTTPS :443 AWS · mcp.gold-dust-lab.de nginx :443 TLS · rate-limit reverse proxy MCP server :8000 FastMCP · 12 tools verifies Duo JWT ask_gold_dust_lab Ollama · qwen2.5:3b :11434 · local brain
Claude path (tools 1–8) qwen branch (tool 9) 🔐 Every /mcp call needs a valid Cisco Duo JWT. Dashed = your AWS box; the shop’s data lives inside the MCP server.

The twelve tools

Talk to the shop in plain language — these are the calls that run behind the scenes.

#ToolAsk it something like…
1get_company_info“What does Gold Dust Lab sell, and when are you open?”
2search_products“What Fender guitars do you have?”
3get_product“Tell me about GDL-R001 — Dark Side of the Moon.”
4search_knowledge_base“What’s your return policy for records?”
5get_order“What’s the status of order GDL-1042?”
6list_customer_orders“What has alice.turner@example.com ordered?”
7create_support_ticket“Open a ticket for bob.keeler@example.com — buzzing fret.”
8request_quote“Quote me a Stratocaster and two Dark Side of the Moon.”
9list_tickets“Does bob.keeler@example.com have any open tickets?”
10get_ticket“What’s in ticket GDL-T5001?”
11get_quote“Show me quote GDL-Q9005.”
12ask_gold_dust_lab“Ask the local assistant: when did order GDL-1042 ship?”runs qwen

Where qwen runs

The local model only takes over on one path. Everywhere else, Claude is the brain and the tools just hand back data.

What you doWho does the thinkingqwen?
Ask Claude chat / Claude Code anything → it uses tools 1–8Claude — in the cloudNO
Call tool 9, ask_gold_dust_labqwen2.5:3b — on your AWS boxYES

Rule of thumb: qwen only runs through ask_gold_dust_lab — which keeps the reasoning and data on your box. Anything else is answered by Claude.