MCPs in My Homelab: Turning Codex Into a Real Ops + Dev Copilot

I spent a couple sessions wiring up MCP (Model Context Protocol) servers in my homelab so Codex can stop guessing and start verifying: live Home Assistant entities, real Portainer containers, Proxmox nodes, GitHub issues, and even WordPress drafts with media.

The architecture (high level)

MCPs we implemented

MCP	What it gives Codex	Runs where
openaiDeveloperDocs	Authoritative OpenAI API / Codex docs lookups	Hosted (OpenAI)
github	Issues/PRs/repo metadata (real triage + cleanup)	Hosted (GitHub)
homeassistant	Entity + service validation, live state lookups	docker_10 (Home Assistant)
portainer	List/manage environments, stacks, containers	codex_appliance (docker_69)
proxmox	Inspect/manage nodes, VMs, LXCs	codex_appliance (docker_69)
wordpress	Draft-only posts + media uploads	codex_appliance (docker_69)

Why this makes Codex better (real examples)

1) GitHub issue hygiene: list open issues, spot quick wins, close stale items

Instead of guessing what’s worth doing, Codex can query GitHub issues directly and group them into:

• Code-only / MCP-assisted (good targets)
• Hardware / house work (not code-only)
• Stale / no longer relevant (close with a note)

Example prompt (what I actually ask Codex):

Show open issues in CCOSTAN/Home-AssistantConfig.
Which are code-only vs. hardware/house?
Which should we close as no longer needed?

From the current open list, some code-only / MCP-assisted candidates include:

• #1539 Create Portainer dashboard + automations (now much easier with Portainer MCP).
• #1517 Create Proxmox dashboard (now much easier with Proxmox MCP).
• #1547 Circuit Panel app improvements (service/app work).

And some that are not code-only (blocked on real-world work):

• #1548 Replace Nest Protects (hardware lifecycle + install).
• #1518 Backup internet (hardware + network changes).

2) Home Assistant safety: validate entities/services before changing YAML

Before touching automations/scripts, Codex can query Home Assistant to confirm entity IDs exist and services are valid. This reduces typos and “it failed because the name was wrong” loops.

3) Infra as context: Portainer + Proxmox MCPs

• Confirm where a service runs and what it’s called (environments, stacks, containers).
• Inspect VM/LXC inventory and node status quickly.
• Turn “go click around” ops work into a single verified conversation.

4) Blogging workflow: WordPress drafts + media uploads

The WordPress MCP is draft-only on purpose: it lets me iterate safely. It can upload images and set a featured image, but I also prepend the hero image inline because some themes don’t display featured images consistently on the front-end.

Snippets from AGENTS.md (how we keep Codex on the rails)

## MCP Servers (This Machine)
- openaiDeveloperDocs: OpenAI/Codex docs and reference.
- github: issues/PR/repo metadata.
- homeassistant: entity/service validation and state lookups.
- portainer: Docker via Portainer (runs in codex_appliance).
- proxmox: VM/LXC management (runs in codex_appliance).
- wordpress: draft-only posts/media (runs in codex_appliance).

MCP placement: if an MCP needs installed dependencies,
host it in docker_69/codex_appliance and expose it to the IDE via SSH stdio.

What I want to improve next

• Better visuals by default: Gutenberg blocks, tables, callouts, and an inline hero image every time.
• Auto-summarize issue lists: a repeatable “triage” prompt that labels quick wins and stale items.
• Media workflow: generate hero + in-post diagrams, upload, set alt text, and insert at the right spot automatically.

Draft generated with help from Codex + MCP servers running in my homelab.