Skip to content

FAQ

No. Reads happen freely, but every change is proposed, then approved — nothing is written until a person signs in and approves it. See Update Fishbowl.

Reads run as a least-privilege, read-only database user that can’t write, on top of a SQL guard that blocks sensitive columns and tables for everyone and suppresses personal fields by default. Writes are approved by a person and run as that person. See MySQL access.

Each connection acts as one Fishbowl user and can do only what that user’s Fishbowl rights and the connection’s scopes allow. An administrator grants access per person and can revoke it at any time.

Only ping. With no scopes, a token or connection can confirm it’s connected and nothing more.

Does the AI write SQL against my database?

Section titled “Does the AI write SQL against my database?”

For admins, yes — sql_query composes a read-only SELECT, checked by the guard. For everyone else, no: the raw-SQL tools are admin-only, and operators read through saved queries instead.

Claude (claude.ai, Desktop, mobile), ChatGPT, Claude Code, opencode, and VS Code / Cursor / Gemini CLI. Claude and ChatGPT sign in by OAuth; the others use a minted token (Claude Code can do either). See Connecting.

Yes. The client, the model, and its reasoning settings all affect how accurate and complete the answers are. A stronger model handles vague requests better, but being specific — and telling the model to ask you to clarify when unsure — gives the best results.

A change set runs its rows in order and stops at the first failure, so a bad row never leaves a half-finished change. Fix that row and re-run.

Reports, pages, and dashboards are versioned — an administrator can roll one back to an earlier version from the Reporting tab.

Revoke it on Connect ▸ Tokens. In an emergency, Revoke all stops every token (or every OAuth connection) at once.