How a project works
Draft and production addresses, numbered versions and rollback, databases, secrets and how vbtnl keeps production data away from the AI.
A project on vbtnl is one app: a site, a bot, an API or a scheduled job. The code lives on vbtnl servers. Your AI agent edits it there through MCP tools, so nothing has to be copied from your computer.
Draft and production
Every project has two HTTPS addresses:
| Address | What it is | |
|---|---|---|
| Draft | yourapp-dev.vbtnl.site | The working copy. The agent edits files here and the app restarts about a second after each change. |
| Production | yourapp.vbtnl.site | The published app your users open. It changes only when you publish. |
Unfinished changes stay in the draft until you say "ship it". Paid plans can put production on your own domain.
Versions and rollback
Each publish takes a snapshot of the draft files, builds it and switches production to it. The snapshot is saved as a numbered version: v1, v2, v3. Asking the agent to roll back switches production to an earlier version in one step, with no rebuild.
If a publish fails, production keeps running the previous version.
What can run
- Python 3.12, for example FastAPI. Dependencies go in
requirements.txt. - Node.js 22. Dependencies go in
package.json. - Static sites: plain HTML, CSS and JavaScript files.
Apps are served over HTTPS. vbtnl does not run GPU workloads or services on custom ports. Settings live in one file, vbtnl.json.
Database
Ask the agent to save data and it adds a Postgres 16 database: one for the draft and a separate one for production. The app reads the address from the DATABASE_URL environment variable.
Production has no persistent disk. Anything the app must keep goes in the database; files written in production are lost on restart.
Secrets
API keys and passwords are stored as secrets and reach the app as environment variables. They are never written into the project files. Tell the agent, for example, "set the secret TG_TOKEN", and give the value.
Scheduled and long tasks
A project can have up to 10 schedules in standard cron format, in UTC: a report every Monday at 9:00, a reminder, a data pull. One-off long tasks run in the background for up to 6 hours.
Your users' data and the AI
Everything a tool returns to the agent is sent to your AI provider. So by default production is closed to the agent:
- it cannot query the production database;
- production logs reach it masked;
- it sees the structure of the production database, not the values.
The agent works with draft data instead. Only the account owner can allow more, in the account settings.