Disclosure first: I help maintain this. MIT licensed, no paid tier, no enterprise edition.

The sysadmin problem it addresses: giving people read access to a database without giving them a path to it. The usual answers are a jump host, a VPN plus a desktop client on every machine, or a port you promised yourself you would close later. Each one adds an account to deprovision when somebody leaves.

LibreDB Studio runs as one container inside the network the database already lives in, and people reach it in a browser. Nothing per workstation, nothing exposed outward.

What makes it relevant to this community specifically:

  • OIDC single sign-on and role based access control in the free build. Hook it to Keycloak, Entra, Authentik or whatever you already run, and offboarding is one account in one place rather than N desktop installs.
  • An audit trail in the admin console. Sign-ins, permission denials, maintenance actions, object edits, and agent and MCP operations, each with the user behind it, plus the signed-in admin's own query history. Exports as CSV or JSON. To be precise about scope: this is not a transcript of every statement every user runs in the editor.
  • Column level masking. Columns the app reads as personal data stay masked until someone with the right role unmasks them.
  • The optional AI agent runs read-only, enforced by the database rather than by a parser. A read-only transaction on PostgreSQL, PRAGMA query_only re-asserted per statement on SQLite, a READ_ONLY handle on DuckDB, and on SQL Server a session principal verified at open to be unable to write. Writes and DDL are refused before the database is reached. That pipeline is the agent's alone; statements you type in the editor go straight to the provider, so for human users the real boundary is still the grant on the connection.
  • 42 drivers reaching 69 engines, so one tool covers the Postgres, the MySQL, the Redis and the Mongo instead of four.

Runs as a Docker image, a Helm chart or a Kubernetes operator, and there are deb and rpm packages if you would rather not run a container.

docker run -d -p 3000:3000 ghcr.io/libredb/libredb-studio:latest

Honest limits: Redis is single node only for now, no Cluster, no Sentinel, no TLS. Druid, Elasticsearch, OpenSearch and Trino take host and port rather than a connection string. Cassandra reports no row count or table size, because the numbers it publishes are estimates that are often wrong, so it shows nothing instead.

Repo: https://github.com/libredb/libredb-studio Demo: https://trial.libredb.org/ (admin@libredb.org / Admin!2026)

Posted this in c/selfhosted earlier today from the homelab angle; this one is written for the people who have to answer for the access list.

AI use, since it comes up: of 2,508 commits in the public history, 31 carry Co-Authored-By: Claude, 8 say "Generated with Claude Code" and 35 mention Copilot, so AI assisted on part of the implementation and review while humans did the bulk. CONTRIBUTING.md states the project's policy.

Edit: tightened the audit and read-only bullets. The first version said the audit trail covers every statement and that read-only is grant-enforced in general; both were broader than what the project's own docs claim, and the text above now matches them.

 

I help maintain this, so flagging it as a promo post up front. MIT licensed, no paid tier, nothing held back.

The problem it solves for a homelab: you spin up Postgres in a container, then you want to look inside it. So you either open a port, dig an SSH tunnel, or install DBeaver on every machine you might touch it from. On a phone you have nothing at all.

LibreDB Studio goes the other way. It is one container you run beside the database, on the same Docker network or in the same cluster, and you open it in a browser. The database keeps no port facing the outside. Nothing is installed per person.

docker run -d -p 3000:3000 ghcr.io/libredb/libredb-studio:latest

What is actually in it:

  • 42 drivers of its own, reaching 69 engines because another 27 speak one of those wire protocols. Postgres, MySQL, MariaDB, SQLite, MongoDB, Redis, Valkey, ClickHouse, Elasticsearch, InfluxDB, Kafka, Neo4j and the rest share one sidebar.
  • OIDC single sign-on, role based access control, a query audit trail and column level masking. These are in the free build, not an enterprise tier. If you run a homelab with family or friends on it, that is the part that matters.
  • Monaco editor with schema aware completion, visual EXPLAIN, ER diagrams drawn from real foreign keys, schema diff that outputs migration SQL.
  • An optional read-only agent that runs against a model you configure, including a local Ollama, so the schema does not leave your network. It is off unless you turn it on.

Install paths: Docker image, Helm chart, Kubernetes operator, deb and rpm, Snap, Homebrew, winget, or one-click templates on CasaOS, Unraid, YunoHost, Runtipi, CapRover, Portainer and Dokploy.

Known limits, so nobody wastes an evening: Redis is single node only, no Cluster, no Sentinel, no TLS yet. Druid, Elasticsearch, OpenSearch and Trino are configured by host and port rather than a connection string. Druid, Elasticsearch and OpenSearch have no UPDATE and no CREATE TABLE in their own SQL at all, so those controls are reported as unsupported instead of failing when used. Cassandra publishes no row count or size that is true, so the object browser shows neither rather than a wrong number.

Repo: https://github.com/libredb/libredb-studio Demo: https://trial.libredb.org/ (admin@libredb.org / Admin!2026)


AI declaration, per the community rules:

  • Implementation (production code): Assisted - AI acted on part of some tasks, humans did the bulk. Measured on the public history: of 2,508 commits, 31 carry Co-Authored-By: Claude, 8 say "Generated with Claude Code", and 35 mention Copilot. That is under 3% of commits, and a trailer is a floor rather than a full count.
  • Review: Assisted - Copilot appears in pull request review alongside human reviewers.

The project's CONTRIBUTING.md states the policy in its own words: using an AI tool to help write a patch is fine, but you must have used the product, reproduced the problem yourself, and stand behind what you submit. Issues and pull requests opened by an autonomous agent are closed without review.