homelab-recommender
A homelab hardware recommendation workflow that runs as a Claude Code plugin and can also be used directly from Codex. It considers DIY builds, prebuilts (Mac mini, Mac Studio, NUC, Beelink, Minisforum), used SFF workstations, and SBCs (Raspberry Pi, Rock 5, Jetson). It fetches live prices and models running costs.
It works especially well for choosing hardware for local AI inference, where tradeoffs like memory bandwidth, VRAM, power draw, noise, and upgrade path matter as much as raw CPU/GPU specs.
Prerequisite: Playwright MCP (recommended)
Live prices from major retailers (Mwave, Scorptec, Amazon, Apple, etc.) require a headless browser. Without it, the plugin falls back to lower-quality price sources. Set this up before installing the plugin.
Add to your Claude Code MCP config:
claude mcp add playwright -- npx @playwright/mcp@latest
Or add manually to ~/.claude.json:
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest"]
}
}
}
Then restart Claude Code.
Install In Claude Code
/plugin marketplace add koustubh25/homelab-recommender
/plugin install homelab-recommender@homelab-recommender
Use In Claude Code
/homelab-recommend
Use In Codex
This repo can also be run directly in Codex without extra plugin packaging.
Open the repo in Codex and ask it to run the homelab recommender workflow from this repository.
Example prompt:
Run the homelab recommender.
If you want live retailer pricing in Codex, add Playwright MCP there as well:
codex mcp add playwright -- npx @playwright/mcp@latest
Then restart Codex. Without Playwright MCP, the workflow can still run, but the pricing stage will fall back to weaker sources.
Runtime Model
- Claude Code uses the packaged plugin under
.claude-plugin/.
- Codex uses the repo directly and follows the same workflow prompts in the repository.
commands/, agents/, lib/, and test-fixtures/ are shared across both runtimes.homelab-run/ is the artifact directory in both runtimes.
Validate Contracts
python3 scripts/validate_contracts.py
This checks the documented JSON handoff shapes across the agents, including DIY, prebuilt, and hybrid candidates.
It also validates a completed golden run fixture under test-fixtures/golden-run/, including constraint-analysis.json, compatibility-report.json, and PLAN.md.
The workflow will:
- Ask you a few questions about what you want the machine to do
- Check your requirements for impossible combinations
- Design 1–3 candidate builds
- Pause so you can react before expensive price scraping
- Fetch live prices from region-appropriate retailers
- Model running costs in your local currency
- Verify compatibility
- Write a
PLAN.md with parts, prices, buy links, and risks
You can change requirements at any checkpoint and the plugin will re-run only the affected stages.
Agents
The workflow is a pipeline of seven specialized agents:
- intake-advisor — Asks one focused question at a time to build a structured requirements profile. Supports patch mode for changing requirements mid-flow.
- constraint-analyzer — Catches impossible requirement intersections (e.g. "ARM + Talos + modular + under $2k") before any build design happens.
- architect — Designs 1–3 candidate builds across DIY, prebuilt, mini PC, SBC, and hybrid types. Reasons from requirements, not from patterns.
- price-scraper — Fetches live prices from region-appropriate retailers using Playwright (preferred), WebFetch, or WebSearch. Reports min/median/max with source quality tiers.
- energy-modeler — Computes idle and load power draw, converts to monthly/yearly cost in local currency, and identifies energy-saving levers.
- compatibility-checker — Verifies socket/RAM/PSU/case fit, BIOS support, OS/distro compatibility, and SKU drift between architect and scraped prices.
- plan-writer — Renders the final
PLAN.md with parts tables, buy links, energy costs, and risks. Pure synthesis — no new decisions.
Region
Defaults to Australia. Other regions are supported via lib/retailers.json and lib/electricity-rates.json — contributions welcome.
