Everything the Agent Can Do Is a Command

Somewhere under every AI agent there is an API doing the actual work. Most platforms keep it private: you see the chat, and you trust that what happens behind it is what you'd have done.

Ours is public. Everything SellerClaw's agents do to run a store is a command, and those commands ship as an open-source command-line client you can install and drive yourself:

pip install sellerclaw-cli
sellerclaw auth login
sellerclaw orders list --status new

SellerClaw itself is a team of AI agents connected to the stores you sell on, plus suppliers and ad accounts, and for most people the chat is the whole product. The CLI is the second door into the same account — for the day you want to script something, audit something, or plug your store into a tool of your own.

Built to be driven by scripts — and by other AIs

The CLI behaves the way automation needs, not the way demos do. Success is JSON on stdout; failure is JSON on stderr with a stable exit code; your token never appears in a command line. Discovery is built in — sellerclaw groups, describe, guide — so a script, or an LLM holding the manual, can find the right call without external docs.

The commands mirror the work, not the marketing: list unshipped orders, publish a listing, sync stock, pull 90 days of analytics, check what an ad campaign spent. If the agent team can do it in a task, there's a command for it here — it's the same API.

Your Claude can run your store

The same package doubles as an MCP server, so Claude Desktop, Claude Code, ChatGPT or any MCP client can drive SellerClaw natively. Instead of dropping hundreds of commands on the model at once, it hands over four: read the manual, see what exists, look up exactly what one command needs, run it. The AI composes the rest itself, the way our own agents do.

The row under the cards: one connection covers Claude in the browser, in Claude Desktop and in Claude Code.

Setup is one sign-in, in a browser, with no API key to find or paste; nothing has to be installed at all if you'd rather paste one address into the assistant you already use. After that, "which of my listings went out of stock this week?" is a question you can ask your own Claude, against your real store.

The gates travel with the account

An open surface would be a liability if it bypassed the rules, so it doesn't: the approval gates live on the server, not in the chat window. Money and outbound email wait for the owner's yes no matter which client asks — the web chat, a cron script, or Claude. A script can't spend what you haven't approved.

The rest of the honesty is structural. The CLI is a client to your account — same permissions, same activity trail as everything else. It's MIT-licensed and on GitHub, so what it sends is inspectable, and it's versioned on PyPI like any other dependency in your stack.

What changes for you

Delegation stops requiring faith. You can hand the routine work to the agent team and still, any afternoon, open a terminal and look at exactly the levers they're pulling — or pull one yourself, or let your own Claude pull it.

If you never install it, nothing is lost; the chat needs no setup at all. But if you're the kind of seller who scripts, start with pip install sellerclaw-cli, sign in, and ask your store a question from the terminal. The same free credits cover the account, whichever door you use.