FAQ

Can't the agent just read the .env afterwards?

It can. A .env on disk is readable by whatever can read files, the agent included, and offprompt does not watch what the agent reads. What offprompt does is keep the value out of the conversation on the way in, and give the agent no reason to look: every result lists the names of all the keys the file holds, and tells the agent that opening the file would put the values into the transcript.

This was learned the hard way. In one cloud session an agent opened the file "to double-check it looks complete", and every value went into its context. The instructions and results have said so plainly ever since. See the threat model.

Why not paste the key into the chat?

Whatever you paste goes into the model's context, the transcript on disk and the provider's logs, and stays there. With offprompt the only thing in any of them is the key's name.

Does offprompt send my values anywhere?

Only into the file. offprompt runs where your agent runs, as a process the agent starts, and its page makes no outbound requests. It sends nothing to us, and there is no offprompt server to send to. From a cloud sandbox the page reaches you through a Cloudflare tunnel, and your browser encrypts the values before they enter it.

Does offprompt keep my values?

No. A request lives in the server's memory while it is open. Once the values are written, the file is the only place they are.

Does offprompt take memory while my agent runs?

Very little until the agent asks for a value. Installed for you, as a plugin, offprompt starts as a small launcher of about 3 MB, which answers the agent's opening questions itself. The server behind it, about 80 MB on Node, starts when the agent first calls one of its tools, and stops again once it has had nothing to do for five minutes. On Windows, and in a project install, the server runs for the whole session, as it did before.

Why does the agent tell me four emoji?

That is the fingerprint. The same four on your page and in the chat mean the file holds exactly what you typed. The agent says them as soon as the write lands, while your page is still open, and offprompt's result opens with them.

Which agents does it work with?

Any that speaks MCP. npx offprompt init sets up Claude Code, Codex, Cursor and Pi in a project, and init -g sets up every agent add-mcp knows, for you. See Agents.

Why does the agent send me a link? I thought offprompt opened the page itself.

On your own machine it does, and a link to click there is the case to distrust. From a cloud sandbox, or a machine with no browser to open, offprompt cannot reach your screen, so the agent passes you the link instead.

What if I paste the wrong key?

The page tells you before anything is written. Each field checks its value as you type, and names the key you pasted when it recognises it: "That's the publishable key. Use the secret key from the same page."

Can it write to Vercel, GitHub or my secrets manager directly?

Not yet: offprompt writes files. To fill the page from a store you already use, see Values you already keep.

Why does my password manager not offer to fill or save these fields?

The fields are text areas, not password inputs, so browsers and password managers leave them alone. That keeps a password manager from saving an API key as a website password.

Is it open source?

Yes, under the MIT licence. The packages bundled into the server and the page's fonts, icons and logos keep their own licences, listed in the notices that travel with the server.