DBX

What it is

DBX is a free, open-source app for looking at and editing databases. Think of it as a lighter alternative to tools like DBeaver, Navicat or TablePlus. It connects to almost any database you're likely to use: Postgres, MySQL, SQLite, MongoDB, Redis, Cloudflare D1 and more than 100 others.

The part that matters for vibe coders is that your coding agent can use it too. DBX includes an MCP server. MCP (Model Context Protocol) is the standard way to give an agent like Claude Code or Cursor a new tool. With it, your agent can list your tables, read their structure and run queries using the connections you've already saved in DBX.

It's Apache-2.0 licensed, built mainly by a developer known as t8y2 with a large, mostly Chinese-speaking community. It's six months old and already has more stars than most database tools on GitHub.

What actually happens when you use it

1. You install the app and add a connection. You paste in your database details, or import saved connections from DBeaver or Navicat. Passwords are stored in your system's password store.
2. You browse and edit. There's a query editor with autocomplete, a spreadsheet-style table view you can edit directly, a diagram of how tables connect, and export to CSV, JSON or Excel.
3. You can ask the built-in AI. Describe what you want in plain English and it writes the SQL. It works with Claude, OpenAI or a local model, using your own key.
4. You connect your agent separately. Add npx @dbx-app/mcp-server to your agent's MCP config. Installing the app doesn't do this for you.
5. You choose what the agent may do. In Settings → MCP you pick which connections the agent can see and one of three modes: Read only, Data read/write, or Full access. Full access lets it drop tables and wipe data.

What people like

Honest note: almost all the discussion happens in GitHub issues (mostly in Chinese) and a Chinese chat group. We found no English hands-on review, no Hacker News discussion (the one submission got no comments) and no real Reddit thread. Online "reviews" mostly repeat the README.

• The maintainers respond fast. When an English-speaking user asked for English issue templates, a maintainer agreed and sent the change the same day. A detailed bug about how the agent tools were labelled was picked up by contributors and closed the same day.
• Users dig in and help each other. When an upgrade broke the agent connection, a user posted a working fix for others while the bug was open.
• The agent safety controls are unusually thorough. On paper, the MCP server blocks broad deletes and table changes unless you choose Full access. It treats tricks like WHERE 1 = 1 as dangerous, and refuses unknown Redis commands.

What people run into

• Updates sometimes break things. After the v0.6.21 update, the agent connection stopped finding database passwords for several users, with a second report of the same thing. One user complained the release seemed to skip basic testing.
• In Claude Code, every query used to need a click. The agent tools didn't say which ones were read-only, so Claude Code asked for approval even for a simple SELECT. That's been closed as fixed, but it shows how fast the agent side is still changing.
• There are a lot of bugs to sort through. The repo has had about 8,150 issues in six months, with 1,245 still open and hundreds of new ones every week. Many are in Chinese, which makes it hard for English speakers to check whether their problem is known.
• Platform quirks. The Docker version can't reach databases on "localhost", which trips people up. A macOS signing problem once blocked local network access. Older Linux systems aren't supported yet.

Before you install

• Writes are not blocked by default. We checked the source code. Until you save a policy in Settings → MCP, the agent server defaults to not read-only, with every saved connection visible. That means it can insert rows and run filtered updates and deletes. Broad deletes, table changes and TRUNCATE stay blocked unless you choose Full access. Set Read only and pick your allowed connections before you connect an agent.
• Never give your agent a production database. Mark live databases as read-only or production in DBX and leave them off the agent's list. Better still, connect only to a local or test copy.
• The AI helper sends data to your AI provider. Your table structure and queries go to whichever model you configure. Use a local model through Ollama if that matters.
• It updates constantly. There were several releases on the day we checked. If the agent connection suddenly fails after an update, check the issues before you start changing settings.
• Finish the security upgrade if you used an older version. Early versions stored database passwords in plain text. Newer ones run an upgrade wizard that encrypts them. Run it, and delete the backup it makes once your connections work.
• Some databases need Java. Snowflake, BigQuery, Neo4j and others work through an add-on that needs a Java runtime, despite the "no Java" pitch.

The short version

DBX is an impressive, fast-moving free database app, and its agent connection is one of the most carefully designed ones out there, on paper. In practice it's six months old, it ships several times a day, and a recent update broke the agent connection. Agents can also change your data unless you switch them to read-only first. If you want your agent to read a local or test database, it's worth setting up carefully. If you only need a stable database viewer, it's fine to try, but keep your current tool for anything important.