The state of homelab tech (2026) (Friends)

Summary of The state of homelab tech (2026) (Friends)

by Changelog Media

2h 2m•January 24, 2026
Listen to Episode

Overview of The state of homelab tech (2026) (Friends)

This episode of Changelog + Friends (hosted by Tim and guest Techno Tim) surveys the 2026 home‑lab landscape: hardware scarcity and rising prices, a huge surge in self‑hosted software and AI-driven tools, and the increasing use of agents/automation to run and maintain homelabs. The conversation mixes practical setups and architecture (TrueNAS, Proxmox, ZFS strategies, NIC/NVMe/bifurcation), hands‑on tooling (Depot, Paperless NGX + Paperless AI/GPT, Proxmox helper scripts, PXM CLI), and higher‑level patterns (ETL/medallion pipelines for documents, agentic databases, “Ralph Wiggum” iterative agent loops).

Key takeaways

  • Single word summary: availability — both scarcity of hardware (CPUs, RAM, HDDs, GPUs) and abundant new software options.
  • Hardware is harder/expensive to source (secondhand market is pricy and constrained by enterprise allocations), so many home labbers are consolidating into “one big box” setups.
  • Software + AI are the growth stories in homelab: self‑hosted ML stacks, RAG (retrieval‑augmented generation) pipelines, vision LLMs for OCR, and agents that can manage infrastructure.
  • Agents and CLI/automation are making previously tedious tasks (VM creation, VLAN troubleshooting, container deployment) far faster and more repeatable.
  • Keep raw inputs (images, scans) and use medallion ETL (bronze/silver/gold) so you can reprocess as models improve.
  • Observability at home (Prometheus + Grafana) has become worth the effort because LLMs help automate setup and scraping.

Topics discussed

Hardware availability & trends

  • Shortages/pricing spikes: server motherboards, DDR5 RAM, HDDs, GPUs (30xx/40xx/50xx markets referenced).
  • Common home‑lab response: consolidate into a powerful single box (NAS + compute + GPU + virtualization) rather than many small nodes.
  • Practical hardware notes: mirrored/striped vdevs in ZFS for expandability, special vdevs (NVMe) for metadata + small files, ARC (RAM) tiering, NVMe via PCIe bifurcation, use of Intel Optane where available, ECC debate (recommended for peace of mind but tradeoffs exist).

Software, AI, and self-hosting

  • Explosion of self‑hosted apps: more container images and AI‑focused services (paperless/document tooling, local model runners).
  • Vision LLMs improve OCR/metadata extraction (Paperless NGX + Paperless AI/GPT example). Models can dramatically raise OCR accuracy for serial numbers, logos, etc.
  • Tools to structure document ingestion: Dockling (IBM open source), Paddle/PaddleOCR; these help break docs into title/footer/tables for better downstream RAG results.
  • Keep original scanned images (bronze layer) so you can retransform with better models later.

Automation, agents, and workflows

  • Agents can be used to audit/fix network configs (example: using Claude to detect/fix VLAN rules on a UniFi UDM Pro via SSH/API).
  • The “Ralph Wiggum” (prompt.md) loop concept: iterative agent loops that run repeatedly until a goal/spec is met. Set iteration limits or budget to control cost.
  • CLI + MCP pattern: build a solid CLI/agent bridge (MCP = model context protocol) so agents can safely perform operations (the guest described PXM CLI for Proxmox).
  • Tests are important: when AI writes code, have it also write tests (proves behavior and documents intent).

Storage & architecture patterns

  • Hybrid ZFS strategy: large spinning disk pools for bulk + special NVMe vdevs for metadata/small files + ample RAM for ARC to achieve NVMe-like responsiveness for many workloads.
  • Pass‑through HBA to TrueNAS VM is a common pattern if you want virtualization + direct drive control.
  • Consider colocation for public‑facing services (bandwidth + uptime) while keeping “home production” on local NAS.

Observability & dev experience

  • Home Grafana + Prometheus is more attainable thanks to AI; Tim and Techno Tim run metrics on many containers and services.
  • Depot.dev approach to build speed: latest generation ARM/AMD CPUs, multiplexed cache strategies, RAM disks for heavy disk operations, and observability to prioritize optimizations.

Notable tools, projects & vendors mentioned

  • depot.dev — fast CI/GitHub Actions runner optimizations (sponsor).
  • Fly.io — container hosting / fast microVMs (sponsor).
  • Notion Agent — personal/team AI assistant inside Notion (sponsor).
  • Paperless NGX + Paperless AI / Paperless GPT — self‑hosted document scanning + AI processing.
  • Dockling (IBM) & Paddle/PaddleOCR — document structure extraction and OCR helpers.
  • Proxmox (VE) — virtualization platform: VMs + LXC; the guest uses Proxmox VMs to host Kubernetes nodes.
  • PXM — a custom CLI Tim built for Proxmox to create VMs programmatically (templates live in userland).
  • Proxmox VE Helper Scripts / Proxmox helper scripts (community) — collection of single‑line installers and “app store” style scripts (community-scripts.github.io).
  • “Ralph Wiggum” loop / prompt.md pattern — iterative agent orchestration approach (emerging concept).
  • TigerData / TaggerData (described as agentic Postgres) — single engine combining vector + keyword search + agent hooks + zero‑copy forks (described in episode sponsor spot).
  • Olama / Open Web UI — local model runtimes and UIs for running models locally.
  • TrueNAS (IXsystems) — NAS + container support (recently moved away from in‑cluster Kubernetes to container-based approach per discussion).

(If you research these further, search the project names as spoken in the episode — some company/product names were spoken in rapid conversation and may vary.)

Actionable recommendations (what to try next)

  • If hardware is constrained: consolidate services onto a powerful NAS box (ZFS pool + NVMe special vdev + GPU) rather than many small nodes.
  • Try Paperless NGX + a vision LLM (Paperless AI/GPT or feed images to your own RAG pipeline) for better OCR and metadata extraction. Use Dockling / PaddleOCR to break complex docs/tables into structured data (Markdown or JSON).
  • Build a medallion ETL pipeline for scanned documents: bronze (raw images) → silver (OCR + cleaned text / structured metadata) → gold (production DB / RAG index). Keep raw images for future reprocessing.
  • Experiment with agents for low‑risk tasks: e.g., ask Claude or an open agent to audit VLAN/Firewall configs or to provision VMs from a CLI. Start with sandboxed changes and enforce iteration budgets.
  • Add observability: use Prometheus + Grafana and have agents help set up scraping/alerts. It’s now much easier and pays off quickly.
  • Explore Proxmox helper scripts for quick LXC/VM app installs (Home Assistant, Pi‑Hole, Grafana, etc.).
  • If you build automation tooling, keep templates and installation/config steps in userland (repo of templates) rather than recompiling binaries for every change.
  • When using AI to generate code or infra changes, also generate tests and idempotent scripts.

Quick checklist / practical snippets

  • VLAN segmentation minimum: trusted, kids, IoT, guests (plus cameras or networking gear VLANs if desired).
  • ZFS basics: mirrored pairs (expand by pairs), special VDEV for metadata + small files, ARC (RAM) for hot cache. Mirror NVMe devices for special vdevs. Back up—special vdev failure can be catastrophic.
  • Multi‑NVMe on one slot: use PCIe bifurcation or an adapter card; ensure motherboard bifurcation support.
  • Automation pattern: CLI/API (Proxmox) ←→ MCP server/skill layer ←→ LLM/agent (human ↔ LLM via prompt.md).
  • Document ingestion: scan → store raw images → run OCR/vision model → structure with Dockling/paddle → index embeddings for RAG.

Notable quotes / soundbites

  • “If I could sum everything up in one word, it'd be availability.” — on hardware & software availability.
  • “This year is the year for self‑hosted software.” — prediction: software explosion will offset hardware scarcity.
  • “Treat containers like apps on your phone.” — practical mindset for trying self‑hosted workloads quickly.
  • “Keep the original image — bronze — so you can reprocess as models improve.” — ETL / medallion advice.

Where to learn more / links mentioned

  • depot.dev (fast build runners)
  • fly.io (hosting / fast machines)
  • Notion.com/changelog (Notion Agent sponsor link)
  • Proxmox VE Helper Scripts / community‑scripts (search “Proxmox VE helper scripts” or community-scripts.github.io)
  • Paperless NGX / Paperless AI / Paperless GPT (search project names)
  • Dockling (IBM) / PaddleOCR (Paddle) — document parsing and OCR helpers

Final note: if you’re short on hardware, invest time now in software, automation, and pipelines. Many of the most exciting homelab wins in 2026 will come from smarter software + agents, not just bigger racks.