Using AddisDB with a coding agent

A coding agent can work with your databases through AddisDB. It can read schemas, run read-only queries and put what it found in front of you in the app. Anything that would change data, a schema or a server waits for you to approve it in AddisDB. It works with Claude Code, Codex, Cursor, VS Code and Windsurf.

Setting it up

Paste this one line into your agent:

Set up AddisDB for me: https://www.addisdb.com/agents.md

The agent reads the setup page, tells you what it will install, and asks before it does. It installs the addisdb command line, plus the desktop app and the editor extension where they apply, checks every download against a published checksum, and registers AddisDB with your agent. No administrator rights are needed.

If you already have the addisdb command line, you can register an agent yourself. Use claude-code, codex, cursor, vscode or windsurf:

addisdb agent setup --agent claude-code
addisdb agent status

The second command shows what is registered. ‘addisdb agent remove’ takes it out again. You may need to restart your agent before it sees AddisDB.

What an agent can do

• Read: list your saved connections, read tables, columns and keys, run read-only SQL, get a query plan without running the query, watch live sessions and server metrics, run one query across several databases and compare the answers, and preview the risks in a migration.
• Show: open AddisDB on a view, put a query or a table in a tab, focus the ER diagram on the tables it is talking about, and load a migration for you to read. A query it shows you runs only if it is a read.
• Change, once you approve: run SQL that changes data or schema, drop a table (a safety copy is kept, and the drop can be undone), import a CSV file, export, back up, generate and clear mock data, make a disposable copy of a database to test on, stop a running query, run a workflow, start or stop a local server, and create a database.

An agent can show things in the desktop app, in VS Code or in the terminal. The desktop app has every view. The other two have fewer, and say so.

You approve every change

When an agent asks for a change, AddisDB opens a dialog that names the agent, what it wants to do, the connection and its environment, with the SQL and the risks in front of you. Prod gets a clear warning. You can approve it, deny it, or deny it and stop the agent. Deny is the button in focus, so pressing Enter refuses. A request you do not answer is refused after 15 minutes.

For reads, and for changes off prod that drop nothing, you can choose not to be asked again. That covers one kind of action on one connection, only while that agent stays connected. It is never saved.

Seeing and stopping an agent

While an agent is connected, AddisDB shows a banner with its name, the last thing it did and the rules in force, along with an activity log and a Stop button. Stop abandons whatever the agent was running and refuses everything it sends afterwards.

Setting the rules

Open Settings and choose Coding agents. One page decides what every agent may do:

• Switch coding agents off entirely.
• Ask me before an agent reads: Never, On prod, or Always.
• Let agents ask to make changes. Switch this off and changes are refused outright.

A rule applies from the next thing an agent tries.

What an agent never sees

• Passwords, tokens and connection strings. An agent refers to a connection by its name. If one is missing, it opens the Add Connection form and you fill it in.
• Values in columns you locked in AI Privacy. They come back as [masked]. If a query uses a locked column anywhere, every value in that result is withheld, and a tool that cannot mask what it returns refuses instead.
• Anything in a database with AI switched off. Its queries are refused.
• More than it needs. Results stop at 200 rows unless the agent asks for more, 1,000 at most, and long values are clipped.

See “Keeping data away from AI” for how to lock a column.

Mock data without an AI key

Generating mock data and the AI steps of a workflow need a model partway through. If you have an AI provider set up in AddisDB, it answers. If you do not, your coding agent answers instead, so you do not need a key for this.

What it costs

Agent features work on every plan. Databases on your own machine are free, and remote ones follow the usual plans. See “Do I need an account?”.