Ghost Holding is one household running as an engineered system. This is a working inventory of what's live — autonomous agents, automations, private infrastructure, and things we've physically fabricated. Nothing here is a mockup. It all runs.
A resident agent with her own inbox, phone number, and Notion workspace. She operates under a tiered-authority model, self-funds through a savings ledger, and reports up without being asked. The rest of the lab exists to give her hands.
Bootstraps from a canonical instruction set, resolves the caller across channels, loads memory, and clears a live flag dashboard on every wake. Acts on reversible Tier-A tasks alone; escalates money, medical, and legal.
SMS via Twilio, email inbox, iPhone Shortcut, Stream Deck buttons, an iMessage bridge, and a live dashboard. One agent, met wherever the household already is.
Every negotiated bill, forced price-match, and killed subscription is logged to a Savings Ledger; every operational cost to a Cost Ledger. Target: monthly earn ≥ monthly cost, three months running.
Four authority tiers from reversible-and-cheap to irreversible-and-founder-only. Sensitive data lives in a vault; the site only ever stores pointers, never raw secrets.
Stacy isn't alone. A small cluster of agents and machines share the load — each with a job, running on a private Tailscale mesh across a Raspberry Pi fleet and a always-on node.
A weekly agent that reads the household's activity, treats each log as a specimen, and surfaces only threshold crossings, repeat patterns, and genuine surprises. Runs on a schedule; stays quiet when nothing crosses the bar.
Dashboard kiosk and hive storage host. Carries the shared drives and serves the always-visible household view.
A portable node on the mesh running local models and a dev toolchain, so work moves with the founder without leaving the private network.
A dedicated always-on machine to become the primary hive node — consolidating storage, background agents, and the NAS role into one resilient box.
The connective tissue: webhooks, edge workers, and scheduled jobs that turn raw inbound signal into filed, classified, actionable state.
Every inbound email lands as a row, gets classified into lanes, and either raises the flag or stays silent. The founder clears; the agent never over-notifies.
An approval bridge and command layer running at the edge — routing requests, holding secrets as Worker bindings, and keeping the control plane off any single machine.
Morning brief, bedtime closeout, and the weekly lab review all fire on their own cadence — no human trigger, and no recurring cost that doesn't earn its keep.
A pointer table that tells every script where each token lives and how to read it — so credentials get reused, not regenerated, and never end up in chat or files.
Every device is a row: CPU, GPU, RAM, spare capacity, IP. Before any purchase, the fleet is checked first — borrow idle capacity before buying new.
Every build, design, and decision is written back to Notion in a form the agent can re-read later — because the system has to remember so the founder doesn't have to.
The lab isn't only software. Designs get modeled parametrically, filed in a CAD library, and printed — then documented so any of them can be re-sliced or remixed later.
Custom-fit parts and kids' pieces designed in OpenSCAD, sliced for the home printers, and archived with full print settings so variants spin off without redoing the work.
A Raspberry Pi fleet and mobile node stitched together over Tailscale — private by default, resilient, and treated as one shared resource pool rather than isolated boxes.
Glanceable, voice-friendly companion apps packaged for smart glasses — surfacing household state hands-free, built from a reusable recipe library.
A household dashboard, Stream Deck control layer, and kiosk view — turning the agent's state into something the whole family can read at a glance.
Happy to talk architecture, agent design, or how a household ends up with its own operating system.
Get in touch