DBHub, from Bytebase, is a popular open-source MCP server for PostgreSQL, MySQL, MariaDB, SQL Server, Oracle and SQLite. It's deliberately minimal, with two tools by default, execute_sql and search_objects. TableStudio's MCP server does the same job from a Mac app and adds what DBHub leaves to other tools. Choose DBHub if you want a free server that runs anywhere, including CI and Linux, or you need Oracle. Choose TableStudio if your agents change data in databases that matter and you want production changes reviewed, dry runs, undo and a record of every conversation.
An MCP server is the start, not the safety net
DBHub gets an agent connected, with a read-only mode, row limits and timeouts. Bytebase says approval flows and audit logging are beyond it. Those are exactly the parts that matter once an agent needs to change a database you care about:
- Approval built in. Production changes are staged as a migration with its reason, SQL and affected rows, applied with Touch ID.
- Dry runs and undo. Writes are tested against the real rows, and every agent change is saved first so it can be undone.
- A record of everything. The Agent Center shows each conversation, and the Audit Log every statement.
- Sandboxes on demand. The agent can copy a database to your Mac and try a migration there first.
- No credentials in config files, and an agent can't widen its own access without your Touch ID.
At a glance
| TableStudio | DBHub | |
|---|---|---|
| What it is | Mac database client with an MCP server and agent plugin | MCP server with a small web workbench |
| Runs on | Mac | Anywhere with Node.js 22.5+ or Docker |
| Price | Plugin free; app $49 once | Free, open source |
| Engines | PostgreSQL, MySQL, MariaDB, SQLite, SQL Server, ClickHouse, CockroachDB, Cloudflare D1 | PostgreSQL, MySQL, MariaDB, SQL Server, Oracle, SQLite |
| Tools | Over 55: schema, queries, dry runs, sandboxes, health and schema checks, index advice, backups | 2 by default, plus opt-in explain_sql, health_check and your own SQL tools |
| How agents connect | Saved connections by name; passwords stay in the Keychain | A DSN or a TOML file holding the connection string |
| Read-only | In the database, per call, 30-second limit | Read-only mode, row limits, query timeout |
| Access per database | Off, Structure only, Read, Read & write | Per configured source |
| Production changes | Staged as a migration for your review, applied with Touch ID | Not covered; Bytebase is suggested for approval flows |
| Dry run with the real rows affected | Yes | No |
| Undo one AI change | Yes | No |
| Record of agent activity | Agent Center, per conversation | Request traces in its workbench |
Blocks psql and migration tools writing to production | Yes, the plugin's shell guard | No |
| A database client for you | Full native Mac app | No |
Where DBHub is stronger
- Runs anywhere. Linux, Windows, CI and remote servers, with stdio or HTTP transport, as an npm package, a Docker image or a Claude Code plugin.
- Free and open source.
- Small and token-efficient. Two default tools keep the agent's context lean, and custom tools let you expose exactly the queries you want.
- Oracle, which TableStudio doesn't support.
Where TableStudio is different
No connection strings in config files. Agents reach your saved connections by name. Passwords stay in the macOS Keychain, so nothing in the repo or an MCP config can leak them.
Production changes wait for you. An agent stages the change, with its reason and SQL, and you apply it with Touch ID. On a remote Postgres database a data change needs a dry run first, and rolls back if the real row count doesn't match.
Every AI change can be undone. Before an agent changes rows, TableStudio saves them, foreign-key cascades included, on every engine, so you can undo that one change without restoring the whole database.
You can see what happened. The Agent Center shows every conversation from every agent: each call, what it read, dry-ran, proposed and changed, and what TableStudio stopped and why.
A real client for you, too. Browse the same databases, review what the agent staged, and fix things yourself in a native Mac app.
Last updated