A new app has no users yet and nobody has its link — so there is nothing to protect. Ship it straight to its live URL and iterate there. Preview comes later, when real people are using it.
Create
somewhere initYour agent creates the project and links the folder. If it needs to sign you in, it asks first — you approve a code in your own browser.
Dev
somewhere devThe app runs on localhost through the same compiler deploy uses, so what you see is what ships. Save a file; the page updates.
Deploy live
somewhere deployOne command and it is on its own live link — database, sign-in, files, email and AI already wired. Then a real-browser check of the live page, from the same terminal.
npx @somewhere-tech/cli init --name my-appcreate the project and link this foldernpx @somewhere-tech/cli devrun it on localhost, compiled the way deploy compiles itnpx @somewhere-tech/cli deployput it on its live linknpx @somewhere-tech/cli browser <live-url>check the live page in a real browsersomewhere preview gives every save a private link with its own copy of your schema, so nothing you try can touch production. somewhere promote ships the exact preview you reviewed — nothing else.
No account yet? A deploy while signed out creates a temporary project with a live link and a claim link. Claim it and it becomes yours.
Copy this. Paste it into your agent.
Those agents run on a machine that persists between sessions, so the CLI there is the whole platform — not a throwaway. Nothing signs you in behind your back: the agent asks, shows you a code, and you approve it in your own browser. After that, the machine stays signed in.
Also works in Claude Code, Codex, Cursor, or any agent with a terminal.
Add Somewhere as an MCP connector.
The same primitives over MCP — deploy, database, files, browser checks — with no terminal. One URL, set up once where your agent lives.
Sign in once. It stays signed in.
ChatGPT Work and Claude Work run your agent on a machine that persists between sessions, so the CLI there is the whole platform, not a throwaway. From that one machine your agent can create the project, run it, deploy it live, and pick it up again tomorrow.
Nothing signs you in behind your back. The agent asks first, then runs somewhere login and shows you an eight-character code with a link. You approve it in your own browser, on whatever device you already use — the machine never needs a browser of its own, and the code is good for five minutes. Only approve a code you can see in your own conversation; one that arrives any other way is someone else’s terminal asking for your account.
Once approved, that machine stays signed in for as long as it keeps being used — every later command and session reuses it — and deploys land in your account, with secrets, email, payments, AI, custom domains and scheduled jobs all available.
Measured, not promised: on 2026-09-02/03, a coding agent given minimal context built and shipped a database-backed app with a real-browser check in fifteen minutes — one deploy, thirteen seconds. Details on request.
Build your product. Not the plumbing.
Every app needs the same plumbing, and every team rebuilds it: sign-in and password resets, data only its owner can read, payments that never double-charge, jobs that retry safely, uploads that stay private. Here it is already built, so your agent starts on your product.
Nothing to stitch together. Nothing to keep in sync.
We handle the plumbing. You build the product.
Everything above, written for the agent itself. The prompt points here already.
Ready when your agent is.
Copy the prompt, paste it, and answer one question: what do you want to build?