How to reverse-engineer a live database

Database already running, ER diagram missing? Connect with a read-only account, pull selected tables onto the canvas, and save a “source of truth” baseline.

Goal: Connect MySQL / PostgreSQL / Oracle / SQL Server and import selected tables.
Prereq: Signed-in project that can configure datasources; prefer a read-only DB user.
Note: The free Demo / public share cannot reach your private DB — self-host or use a signed-in instance.

Steps

1. Open the project → menu → Reverse engineer datasource (empty state may show Reverse from datasource).
2. Select or create a datasource, fill connection details, test connectivity.
3. Pick schemas / tables (for large DBs, filter by prefix or allow-list first).
4. Run the parse; wait for completion.
5. (Strongly recommended) Save a version as the reverse baseline.

What success looks like

• Selected tables appear; FKs show as relations within supported scope.
• Comments / indexes are carried over per dialect capability.
• Connection passwords do not appear in exported model JSON.

Troubleshooting

Symptom	Try
Connection failed	Host/port/firewall; in Compose networks use the service name, not localhost for the host machine
Permission errors	Use a read-only account with dictionary privileges
Too many tables / timeouts	Narrow selection; reverse by schema or prefix batches
Missing indexes or FKs	Read parse hints; some dialect features are skipped, not invented
No reverse entry on Demo	Expected — self-host or a signed-in instance. The public Demo is not a PAT
Agent just called create_version	Open the same version diff, accept or roll back. Pick prompt suggest-erd-version. API 200 is not human approval

Next

• File-only / no JDBC: Import DBML
• Let Cursor suggest a version (you still diff): MCP guide — prompt suggest-erd-version
• Secrets isolation: Security model
• Start here