Homelab AI

A self-hosted environment for local AI models, agents and reproducible development workflows.

2026 · INFRA

All projects

Overview

Homelab AI started as a way to run large language models locally, but gradually grew into a complete development environment for experimenting with models, autonomous agents and isolated project workflows.

The goal is not to reproduce the cloud at home. It is to have a system I can understand, change and rebuild — while keeping experiments isolated and repeatable.

Capabilities

Local models

Multiple language and vision models can be served locally and switched depending on the task.

Agent workflows

AI agents work with code, search and development tools inside controlled project environments.

Reproducible environments

Project VMs are built from small golden templates, keeping development environments isolated, disposable and predictable.

Infrastructure as a project

Configuration, recovery procedures and supporting tooling are documented and versioned alongside the system itself.

Design principles

Local first

Models, tools and development environments run locally whenever it makes practical sense.

Disposable over precious

Environments should be easy to rebuild, replace or reset instead of becoming fragile one-off setups.

Automation with an escape hatch

Repetitive work is automated, but the underlying system remains understandable and accessible when something breaks.

Document the boring parts

Backups, recovery, maintenance and configuration are treated as part of the project, not as afterthoughts.

Technology

HOST
Linux · Btrfs · KVM/QEMU · libvirt
AI
llama.cpp · Hermes Agent · local LLMs
DEV
Debian VMs · CodeGraph · Codex · project scaffolding
INFRA
K3s · Caddy · SearXNG · Grafana · custom telemetry

Current state

The system is in active use and evolves together with the projects built on top of it. Recent work has focused on reproducible VM templates, agent tooling, model routing and automated review workflows.

System architectureOne physical host supports loosely coupled local AI, agent memory, virtualization, development VMs, source control and backup storage.Host: onephysical systemLocal AI: modelruntimeAgent /memory:persistentknowledgeVirtualizationDevelopmentVMs: isolatedprojectsSource control:versioned workBackup:recoverystorage System architectureOne physical host supports loosely coupled local AI, agent memory, virtualization, development VMs, source control and backup storage.Host: onephysical systemLocal AI: modelruntimeAgent /memory:persistentknowledgeVirtualizationDevelopmentVMs: isolatedprojectsSource control:versioned workBackup:recoverystorage

One physical host is the foundation for loosely coupled local AI, agent memory, virtualization, development VMs, source control and backup storage.

One host. Independent layers.
Disposable development environmentsA golden image creates isolated development VMs. Develop, test and review inside the VM, then preserve changes in source control. The VM can be discarded while project state survives.Create freshGolden image:reusabletemplateDevelopmentVMDevelopTest and reviewSource control:keep changesDiscard the VM Disposable development environmentsA golden image creates isolated development VMs. Develop, test and review inside the VM, then preserve changes in source control. The VM can be discarded while project state survives.Create freshGolden image:reusabletemplateDevelopmentVMDevelopTest and reviewSource control:keep changesDiscard the VM

A reusable golden image creates isolated development VMs. Develop, test and review inside the VM, then keep changes in source control. The VM can be discarded while project state survives.

Replace the environment. Keep the work.
Central review flowLightweight hooks in disposable project VMs send work to one persistent central reviewer. Review results return to the project workflow to continue or revise.ResultProject VM:hooks / clientReview: onepersistentcentral serviceProject VM:hooks / clientProjectworkflow:continue /revise Central review flowLightweight hooks in disposable project VMs send work to one persistent central reviewer. Review results return to the project workflow to continue or revise.ResultProject VM:hooks / clientReview: onepersistentcentral serviceProject VM:hooks / clientProjectworkflow:continue /revise

Lightweight hooks in multiple disposable project VMs send work to one persistent central reviewer. Review results return to the project workflow to continue or revise.

Many project environments. One reviewer.
Agent memory and local AIAn agent accesses persistent memory, connected to embeddings for finding context and a modular local model runtime. Models can be swapped without rebuilding memory.Agent: task /contextMemory:persistentknowledgeEmbeddings:find contextLocal AI: modelruntimeSwap models Agent memory and local AIAn agent accesses persistent memory, connected to embeddings for finding context and a modular local model runtime. Models can be swapped without rebuilding memory.Agent: task /contextMemory:persistentknowledgeEmbeddings:find contextLocal AI: modelruntimeSwap models

The agent accesses a separate persistent memory layer, connected to embeddings for finding context and a modular local model runtime. Models can be swapped without rebuilding memory.

Persistent knowledge. Replaceable models.

Built to be changed.

The most useful part of the project is not any single model or tool, but having an environment where replacing one does not require rebuilding everything around it.