How REM Vibe Coder keeps your apps safe
An AI that writes and runs code needs walls around it. The sandbox, the network rules, the permission cards, the redaction and the push filter that REM Vibe Coder puts around every build.
REM Vibe Coder writes code and runs it: it installs packages, starts servers, runs tests, clicks through your app in a browser. That is what makes it useful, and it is also exactly why it has to be contained. Here is what stands between a build and everything else.
One sandbox per project
Every project gets its own container, used only for that project, with:
no administrator rights: it runs as an ordinary user, every Linux capability is dropped, and it cannot gain new privileges;
a read-only system: only the project's workspace and two small scratch areas can be written;
fixed limits: 2 GB of memory, two CPUs and 512 processes, so one build cannot starve another;
its own private network, separate from every other project;
no Docker socket: a build cannot start or touch other containers.
Optionally the sandbox runs under gVisor, a second kernel boundary between the build and the machine.
A sandbox that has been idle for 30 minutes is removed; your code is kept and the sandbox is recreated from it in about a minute the next time you build or open the preview.
A network that only reaches what it needs
Outside a developer's own laptop, a sandbox has no open internet. It reaches the package registries a build needs (npm, PyPI and the like) through an allow-list proxy, and nothing else. If that proxy is missing, the engine refuses to start sandboxes at all: it fails closed, not open.
Your preview is yours
The preview of an app being built is served through Vibe Coder, behind a short-lived ticket bound to you and that project. Nobody else can open your work-in-progress app by guessing an address, and a chat, run or project that is not yours answers "not found", as if it did not exist.
The agent asks before it acts
Some actions are routine parts of every build: creating the app's database, saving its code to its own work branch. Everything sensitive is not. Before the builder deletes something, sends something, schedules something or changes a live system in your name, the build stops on a card ("Change the live app?", "Remove something?", "Use or change your data?") and waits for Allow once, Always allow (this chat) or No.
Two more rules sit behind those cards:
Helper agents cannot ask: a sensitive action inside a helper is simply refused. Only the main agent can ask you.
When REM CloudHub asks for confirmation of a change, the confirmation code is never the agent's to give. Only your Allow sends it, and a stale code is never reused.
Acting as you, never as more
Builds act in REM CloudHub through its tools with your permissions. A tool your role there does not allow is not offered to the builder in the first place. Your own MCP servers are called through REM CloudHub's gateway, so their credentials never enter the sandbox.
Secrets do not leak into the record
Everything a build does is an event in your chat and a span in the operators' telemetry. Before any of it is stored, secrets are stripped out: tokens, keys and passwords in tool output, URLs or error messages are redacted. Your session never goes into a URL; the live stream and test screenshots use short-lived signed links instead.
Only your app leaves
When code leaves Vibe Coder (the git push, the deploy, the zip you download), one filter decides what goes: your app's files only. Vibe Coder's tooling, the sandbox's state, internal helpers and credentials stay behind. The generated README has no credentials in its examples, and REM CloudHub's own push check refuses any URL that carries one.
Sign-in you can trust
There are no Vibe Coder passwords. Everyone signs in through REM Auth; an existing account is linked to a REM identity only when REM Auth has verified the e-mail address, an unverified identity is never more than a member, and access is re-checked every five minutes, so someone switched off in REM Auth is signed out of Vibe Coder at once.
Checks before "done"
Safety is also correctness. An app is called done only when its unit tests and a real browser workflow test pass on the current code and its preview answers. A build that cannot get there after three rounds ends as failed and says why, rather than handing you something broken.
Questions about security? Write to support@remplatforms.com. The developer docs describe the authentication and webhook signatures in detail.
Keep reading
From idea to deployed app: a REM Vibe Coder walkthrough
One habit tracker, from a new REM account to a tested app with its own database - every screen, every check and what each step took.
Build an app by describing it
What happens between the sentence you type into REM Vibe Coder and the app that comes back - and how to write the sentence so you get the app you meant.