Every team I've worked with has the same problem with cloud sandboxes: they take a ticket to get, a week to arrive, and nobody ever deletes them. So I built a factory for them on Oracle Cloud. You describe what you need in a chat, it builds it in your tenancy with Terraform, hands you the links and credentials, and destroys it when its lifetime ends.
Here it is, start to finish, in under three minutes:
What the video shows
- A chat with nothing running.
- Grafana from one sentence. "Deploy Grafana for my team from the public image on port 3000." The assistant plans the container, prices it from Oracle's published price list, and notices the code needs an admin password. It asks for it in a box on the page. The value never reaches the AI. One click, and it's building.
- A migration question. "I have Airflow writing to S3 and Glue building Iceberg tables. Move it to OCI." It maps each piece: S3 to Object Storage, Glue jobs to Spark on Data Flow, the Glue catalog to Data Catalog, and Airflow stays Airflow. Your DAGs run unchanged.
- The repository. Paste a GitHub folder link. It reads the DAG and the Spark job and builds the pipeline: a bucket for raw and gold data, a Data Flow application, a Data Catalog, Airflow with the DAG loaded, and an Autonomous Database for the gold tables.
- The tour. Grafana, signed in with the password you gave. Airflow with the DAG green in four steps. The gold tables served as REST straight away.
- Ask the data. A chat UI over the database, and an MCP endpoint so Claude, Cursor or any agent can discover the tools, pick the tables and run the query.
What it will build
| Ask for | You get |
|---|---|
| A database to explore or to wire into agents | Autonomous Database with Select AI and REST on, a chat UI, an MCP endpoint |
| Your app from a Git folder with a Dockerfile | The image built inside OCI, served on HTTPS |
| An AWS Lambda, or any small handler | An OCI Function. Lambda code runs unchanged, optionally on a schedule |
| A Glue or Spark job | A Data Flow application, optionally writing Iceberg tables, with a query app to read them |
| An Airflow plus Spark pipeline | Airflow with your DAGs, Data Flow, Data Catalog, gold tables in Oracle |
| Queues, NoSQL, Kafka, buckets | OCI Queue, NoSQL tables, a Kafka cluster, Object Storage |
It builds the design you name and nothing you didn't ask for. If you say "no database", there is no database. If you describe code you don't have yet, it writes it.
The parts I'm proud of
Honest pricing. Every proposal comes with a cost table from Oracle's live price list, per part, with the free allowances and the per-use services called out. A paid Autonomous Database is about $500 a month while it exists, so the assistant defaults to the pay-per-use pieces and prices the database as an option you can drop.
It follows your architecture. Paste a mapping and it builds exactly that list. It may say in one sentence that another design would be better. Then it builds yours.
Sandboxes die. Each one has a lifetime of 1 to 30 days and a budget. A reaper destroys it when time is up. Every card has a Destroy button for sooner.
Nothing leaves your tenancy. The install is one Resource Manager stack. The workers, the database that runs the app, the AI calls to OCI Generative AI, all of it stays inside your account. Users sign in to the app with their own login, no OCI account needed.
Install it
One click. It needs a tenancy administrator and a Pay-As-You-Go or paid account, and it takes about 15 minutes.
The Terraform, the prerequisites and the example workloads are on GitHub: ashishsinha1602/oci-sandbox-factory. Read the prerequisites first; they list exactly what the install creates and what it grants.
How it's built
- Terraform on OCI Resource Manager for everything, the install and every sandbox. No local tooling.
- APEX on an Autonomous Database for the app. The assistant runs inside the database and calls OCI Generative AI.
- Container instances as workers. They turn requests into stacks, build images inside OCI with kaniko, and never hold a laptop's credentials.
- Select AI switched on in every database, so any table can be asked in plain English.
- An end-to-end suite that runs inside OCI and builds every sandbox type for real, checks it live, and destroys it.
What's next
RAG and agents with only what Oracle provides: Select AI's vector index crawling a documents bucket, images described as text first so pictures and scans are searchable, Select AI Agents over the documents and the tables, and Oracle's own MCP server.
Try it, break it, and tell me what you'd want it to build next.