The Gateway Papers
Personal AI Infrastructure
Architecture Explainer

Your phone isn't the assistant. It's the remote control.

A new wave of AI assistants (like OpenClaw) splits the system in two: a thin app on your phone, and a gateway — the actual agent — running on hardware you host. Drag the model below, then flip between architectures to see where your data actually lives.

Interactive: where does the agent live?

Drag to orbit · works with touch · use the controls below
Your hardwareAgent runs on
You (local disk)Memory controlled by
Stays homeConversation data
~$5–15/moTypical infra cost

How the split works

In a gateway architecture the phone app does almost nothing intelligent. It captures your text or voice, encrypts it, and forwards it to a gateway process — typically a small server you run on a Raspberry Pi, a home server, a NAS, or a $5/month VPS. The gateway is the real agent: it holds your long-term memory, decides which tools to call (calendar, files, smart home, model APIs), executes them, and streams the answer back.

Worked example: you ask “reschedule my dentist and text Sam.” The phone sends about 1 KB of text. The gateway loads your memory (“dentist = Dr. Rivera, Sam = +351…”), calls a language model API for planning, then calls two tool endpoints. The only party that ever sees your contact book is the machine in your closet.

1 · Interface layer

Phone, watch, laptop — thin clients. Replaceable, stateless, trusted with nothing but the current session.

2 · Gateway (the agent)

Your always-on process: memory store, tool router, model access keys, permission rules. One gateway serves all your devices.

3 · Capability layer

Model APIs and tools the gateway calls on demand. Swappable: change the LLM provider without touching the app.

Self-hosted vs. company cloud

QuestionSelf-hosted gatewayCompany cloud assistant
Who stores chat history & memory?You, on your diskThe vendor's database
Can the vendor change or kill it?No — you run the binaryYes, any time
Model choiceAny API or local modelWhatever they ship
Typical cost~$5–15/mo VPS + pay-per-token$0–20/mo subscription
Setup effort30–60 min, some CLI comfortInstall app, sign in

The trade is honest: self-hosting costs you an hour of setup and a little upkeep, and buys you portability, privacy, and an assistant that can't be discontinued out from under you.

Enjoy this tool? Build your own with Super