The Agent Can Live on Your Hardware

"Self-hosted" usually describes a trade. You get control and lose the product: the polished app becomes a README, and you become the person who babysits a server at midnight.

This is the other kind. One command puts the AI agent that runs your store on a computer you own — a desktop, a home server, a small VPS:

curl -fsSL https://get.sellerclaw.ai/agent.sh | sh

And that's the whole trade: the compute moves, the product doesn't. You still sign in at sellerclaw.ai, still talk to the agent in the same chat, still see the same stores, orders and reports. What changes is where the work physically happens — the browser sessions, the tool calls and the LLM traffic run on your machine instead of ours.

One command, and you can read it first

The script checks the machine, installs Docker if it's missing, pulls a prebuilt image, starts it, and prints a code you confirm in your browser — at which point the agent is paired with your SellerClaw account. Nothing to clone, nothing to build; the runtime, the browser and the tooling all live inside the image.

The first line is the trade: Your own hardware — no hosting credits charged. Below it, the agent's version, status and the one install command.

If piping a URL into sh makes you wince — good instinct. Download the script and read it before running it; the connection protocol and the runtime's contracts are public, and the code is on GitHub. Afterwards, management is four commands: sellerclaw-agent status, logs, update, uninstall.

The requirements are modest and worth knowing up front: Docker-capable hardware with more than 2 GB of RAM, on Linux or macOS.

Nothing on your network is exposed

The pairing is outbound-only. The agent connects out to SellerClaw over HTTPS and pulls its work; there is no public IP to arrange, no port forwarding, no tunnel, nothing on your network listening for the internet.

The runtime itself is confined to its container — it doesn't get your host system, your files or your browser sessions. Flip Settings → Hosting in the web panel and your tasks start running on the machine under your desk; flip it back and they return to the cloud.

Who actually wants this

Three kinds of people, honestly. Sellers who'd rather bring their own compute than pay for managed runtime hosting — a laptop that's on anyway, a small VPS running long-lived automations. Privacy- and residency-minded teams that want the agent's browsing and model traffic on hardware they control. And builders who want to poke at a runtime whose wire contracts are public.

Two limits belong in the decision. This is not a standalone agent: it's designed around a SellerClaw account, and without one there's nothing to drive it. And it's young — working, documented and open for experimentation, with stable contracts but internals that still evolve. If "boring and settled" is the requirement, the cloud runtime is the default for a reason.

The same principle is extending to the models themselves: you can bring your own AI provider keys, or point the agent at your own LLM endpoint, so the thinking runs on accounts — or machines — that are yours too.

What changes for you

Where the agent runs becomes a decision instead of a property of the product. Most people will never move it, and lose nothing. The ones who need it — for cost, for privacy, for the principle of the thing — get it without giving up the panel, the chat or the team.

Either way the account starts the same: sign up, connect a store on the free credits, and the hosting switch will be there when you have a reason to flip it.