A coding agent should not stop when the code works locally. It should stop when the application works on a real domain.
A generated application is not a deployed product. Someone still has to configure infrastructure, create a database, manage secrets, add authentication, build a container, route a domain, issue TLS, read logs, repair failures — and check that the public application actually works.
Coding agents stop at one of three false finish lines:
Same app. Same result. The only thing that changes is how much of it lands on you.
11 stepsall yours
1 sentencethen nothing
put my app online
nothing for you to do
Live. your-app.rightside.host
Five stages. The agent drives all of them; you approve the ones that matter.
Install the Rightside MCP server for Claude Code or Codex and approve the device in your browser. Credentials land in protected local storage — never pasted into the conversation, never returned in tool output.
No new syntax to learn and no console to switch to. The agent already has the repository open; it now also has somewhere to put it.
Rightside creates the project, a dedicated PostgreSQL database, the application identity boundary, protected secrets, an isolated container, a generated domain, routing and an HTTPS endpoint.
Not six dashboards. One call, one release, one URL.
Every failure carries the stage it failed at, a stable error code, redacted context and a suggested fix. The agent changes the application and retries — instead of handing you an opaque log dump.
A successful deployment returns the version, the status and the real HTTPS address. The agent probes the public endpoint before it reports completion — because a returned URL string is not the same thing as a working application.
MCP tools and a CLI designed for Claude Code and Codex — stable exit codes, machine-readable output, retry-safe operations, so a retrying agent cannot create a second project or build a second release.
Per-project credentials, injected into the application runtime only. Data survives redeployment and rollback; rolling back code never rolls back the database.
Email and password handled at the gateway. Your app receives a signed identity header, not a session to manage.
Every deployment is a numbered release. Roll back to any previous healthy version; the active release only changes once the new one passes its health check.
Every failure carries the stage it failed at, a stable error code, redacted context and a suggested fix the agent can act on.
Tokens, passwords and connection strings are stripped before a log reaches the agent — debugging never leaks a secret into a model context window.
Routing and TLS provisioned with the project. The deployment is not reported complete until that address answers.
The agent probes the public endpoint before it reports completion — a returned URL string is not the same thing as a working application.
The control plane and the runtime are separate machines. Containers run non-root with dropped capabilities on internal networks, every project gets its own credentials, and secrets are stripped from every log and error before they leave the platform.