STOP RENTING TWELVE SAAS SEATS. BUY THE ROOM ONCE. RUN THE LAB FOREVER.
Your team already pays for chat, search, a vector store, an IDE plugin, a workflow tool, and a second UI “just in case.” Every prompt leaves the building. Every seat is a recurring invoice. A 12 GB workstation can host the whole motion if someone will compose it, name the network, and refuse to run three GPU engines at once.
That is this lab. Not a slide. A running Docker project named tabby-tavern on tabby-tavern_ai-network. Primary inference is Qwen3.5-9B through llama.cpp CUDA. Chat, RAG, search, MCP, browser VS Code, n8n, and character cards sit on the same bridge. TabbyAPI EXL3 and Ollama stay in an alternate-inference profile so the 4070 does not eat itself.
/v1.diagnostics profile. Socket stays internal.Old Taproot and qwen38-agent containers are stopped. Their YAML is folded in, not dangling as a second project.
Verified up on the lab box after the merge. Host URLs from WSL. Inside Docker, use the service name — never localhost.
Qwen3.5-9B GGUF. Alias Qwen3.5-9B. Containers call http://qwen:8080/v1. This is the primary engine.
Canonical chat UI. OpenAI API to Qwen. SearXNG RAG. One UI — not a second Taproot WebUI on 3001.
Character frontend. PNG cards ship in cards/. Import Seraphina. 401 on root is auth doing its job.
Private metasearch, JSON on. Open WebUI and MCPO search through it. No tracking SaaS.
MCP → OpenAPI. Stack tools plus tavern connectivity. Swagger at /docs.
Browser VS Code. Continue + Tabby extensions. Wired to Qwen /v1 for chat and autocomplete.
Vector store for RAG. Health /healthz 200 on the live lab.
Lightweight embeddings DB. Basecamp volume preserved.
Workflow automation on the same network as the model. 400+ integrations, local.
Extra chat UI pointed at Qwen. Keep Open WebUI as the source of truth.
RAG workspace. Bring docs to the same box that runs the model.
Multi-model frontend. Mongo on the bridge. JWT from your .env.
Full-text search. Health 200 after merge.
Hermes memory backend. Healthy. Password is a placeholder in the public tree.
LibreChat data. Named volume kept from Basecamp.
cd tabby-tavern cp .env.example .env # generate real keys docker compose up -d # 15 default services, Qwen only on GPU # optional: docker compose --profile alternate-inference up -d tabbyapi ollama docker compose --profile diagnostics up -d --build dockroot
/v1 — http://qwen:8080/v1 from another container.Qwen3.5-9B, not the word “qwen.”Jeremy Panasuk — enterprise data / platform background: Kohl's Informatica (PowerCenter Grid/HA, AIX→commodity Cisco/IBM, DQ/MDM POCs, ~$2M software); Dawn, MF Global, BNY, Micron, NFCU, Mercedes, Smith & Nephew. Independent local-AI year from Aug 2024: this Docker lab, decode telemetry, multi-agent prototypes, Hermes on the same box.
"You sent one of the best that Informatica has to offer in Jeremy." — MF Global Infrastructure
Three ways to put UNIFIED on your metal. 30 days email support after deploy.
THE CLOSE. This is not a multi-tenant SaaS. It is a private lab, production-shaped, that Hermes just reconciled onto one compose file. If the requirement is “one network, one Qwen endpoint, chat + RAG + IDE + MCP + character cards on hardware we own,” UNIFIED is the tree that matches the running box.