CJW–08

personal lab

local models · agents · memory

modular systems, not a single chatbot

guide

V.0.1.0

one of the most active areas of the homelab is local ai — not as a toy chat window, but as a modular system of models, gateways, memory, tools, and interfaces.

the goal is a systems-of-systems architecture: agents that can reason, remember, and act against infrastructure with clear boundaries.

interface → agent → models
  1. rabbit r1 / human
  2. hermes agent
  3. litellm gateway
  4. ollama / local models
hermes capability fan-out
  1. hermes
  2. honcho memory
  3. skills / terminal
  4. docker sandbox
  5. higher-reasoning route

stack

1.1

hermes

hermes is the agent control plane in the lab — a place where prompts become tool use, memory lookups, and model routes instead of a single opaque reply.

1.2

litellm

litellm acts as a model gateway: one interface, many backends. agents call logical names like local-fast or atlas-reasoning without caring where the model actually runs.

local-fast · local-code · atlas-general · atlas-reasoning · cloud-premium

1.3

ollama

ollama provides local model execution. qwen and other general-purpose / embedding models run against the lab gpu so everyday inference can stay on-prem.

1.4

memory

honcho-style memory is explored for persistent agent context — infrastructure facts, project history, architecture decisions, troubleshooting notes, and preferences.

infrastructure · projects · architecture decisions · troubleshooting history · recurring tasks · preferences

1.5

rabbit r1 as front end

the rabbit r1 is treated as a physical interface into the personal agent environment — human → device → hermes → memory, tools, models, and home infrastructure.

1.6

agent tooling

conversation is not the end state. the lab direction is agents that can inspect and act — carefully, with sandboxes and clear scope.

agent ambitions

  • —execute terminal commands in a sandbox
  • —query kubernetes and inspect logs
  • —validate services and run health checks
  • —interact with apis
  • —work with git and documentation
  • —generate code and orchestrate workflows

((kicker))

design rule

agents should consume logical capabilities — models, memory, tools — without hard-wiring every call to a single vendor endpoint.

← back to homelab