Hermes skill44 recipesTabby Tavern companion

Local AI Stack Connectivity

A verified playbook for wiring Docker inference engines, chat UIs, search, and MCP gateways on one network. Distilled from the Tabby Tavern lab — not theory.

THE CONNECTION RULE — wire any OpenAI-compatible UI to an engine:

Path is always /v1 → http://<host>:<port>/v1. Never /api/v1. Never put /v1/chat/completions in the base URL.
Host is the in-network service name, not localhost. Inside Docker, localhost is the container you are in. Open WebUI talks to TabbyAPI at http://tabbyapi:5000/v1.
Model is the exact served id (e.g. llama3.1:8b), not the word “ollama”. Ask /v1/models.
If the engine has auth, pass its API key. Do not “fix” a 401 by changing the host.

What’s in this Space

File	What it is
`SKILL.md`	Full Hermes skill: connection rule, discovery, wiring workflow, pitfalls
`references/wiring-recipes.md`	44 recipes — config location, keys, one-line fix, verify step

Pitfalls already paid for

Ollama global 64K context → ~6 GB KV cache → 8B weights pushed to CPU. Use flash-attn + q8_0 KV instead.
Two heavy engines on a 12 GB card: the second CUDA alloc can block forever instead of erroring. Free VRAM first.
TabbyAPI current images dropped EXL2. Point model_name at an EXL3 directory.
Empty MCPO mcpServers: {} crashes the proxy. Keep at least one server.

Install as a Hermes skill

mkdir -p ~/.hermes/skills/mlops/local-ai-stack-connectivity
cp SKILL.md ~/.hermes/skills/mlops/local-ai-stack-connectivity/
cp -r references ~/.hermes/skills/mlops/local-ai-stack-connectivity/

Reference implementation

Tabby Tavern stack v2.0.0 · dockroot-mcp · sell sheet

Skill author: Jeremy Panasuk. License: Apache-2.0. Hermes Agent by Nous Research.