We build an open-source SQL client that a team runs on its own server. One browser tab, eighteen database drivers, MIT licensed. Repo: GitHub - libredb/libredb-studio: One browser tab for PostgreSQL, MySQL, Oracle, SQL Server, MongoDB, Redis, SQLite, Couchbase, ClickHouse, Druid, DuckDB, Turso and more. An open-source SQL IDE with SSO, audit trail and AI-assisted queries MIT licensed, with nothing held back behind an enterprise wall. · GitHub
The reason it exists: most database GUIs are either desktop apps that each person installs and configures separately, or hosted services you point at your production database. We wanted the third option, a web client that lives inside the same network as the data and never asks the database to open a port to the internet.
What it does:
Eighteen drivers: PostgreSQL, MySQL, Oracle, SQL Server, SQLite, libSQL, DuckDB, MongoDB, Redis, Couchbase, ClickHouse, Apache Druid, Elasticsearch, OpenSearch, Apache Cassandra, Trino, Prometheus and Apache Kafka. Twenty-eight more engines speak one of those wire protocols and connect through an existing driver, so the reach is forty-six named engines.
Five of the eighteen are read-only by design: Druid, Elasticsearch, OpenSearch, Prometheus and Kafka. We say that explicitly rather than listing them as fully supported.
Schema browsing, a query editor with schema-aware autocomplete, query history, result grids, schema comparison between two connections, ER diagrams and data profiling.
An agent mode that runs against your own model endpoint, including a local one through Ollama. It drafts SQL and never executes it. Applying a statement is your click.
Deployment is one docker command and the database never needs a public port.
Two honest notes, because a showcase post that only lists features is not much use to anyone:
First, we measured the agent mode instead of claiming it works. Eleven days, thirty-nine open-weight models run locally, 8,199 runs. The uncomfortable result was that most failures were not the models. Of the losses attributed to the model, 75.7 percent came from runs that had already invoked a tool correctly, and the cause was defects in our own server. We published the protocol, the failures and the raw data: [2609.21341] What Stops a Small Language Model From Driving a Database Agent
Second, compatibility claims are easy to inflate. Before writing a number next to the word engines, we ran twenty-eight borrowed engines under the same checks. Not all of them pass every panel, and the per-engine results are in docs/providers rather than hidden behind a marketing number.
Happy to answer questions here. If anyone wants to try it against an engine we have not measured well, that is the feedback we most want.