How AI agents act
An AI agent acts right away as the person who connected it, with their rights, or Rowsafe refuses and says why.
An AI agent connected to Rowsafe acts as the person who connected it, right away, with that person's dashboard rights. Like a doctl or AWS CLI token.
- Rowsafe makes the change at once, or refuses it and nothing changes. Nothing waits for an approval.
- The agent checks with you in chat before anything disruptive, destructive or paid.
- The same checks as your click run, and the safety nets stay on.
- Every change is in the audit log, like [email protected] through their AI agent Claude.
Who an agent acts as
| The agent connects with | It acts as |
|---|---|
| An app signed in with Rowsafe (ChatGPT, Claude, Cursor, the Claude Code and Codex plugins, ...), with Act for you ticked | The person who clicked Allow |
| An API key made in the dashboard (Settings → API keys) | The person who made the key |
The CLI's login (rowsafe login), which a local rowsafe mcp uses | The person who approved the login |
Rowsafe checks that person's role just before each change:
- Owner or admin: runs right away, as them.
- Member: refused, as in the dashboard.
- Left the organization: refused. Their apps are disconnected; their API keys stop working for changes until revoked (API keys).
- Owner-only stays owner-only, like a first payment.
An agent never gets more than its person's rights, and can never change what AI agents may do.
Agents that can't make changes yet
- An API key made by another key (like
rowsafe api-keys create; shown as Made by key ...). Make the key in the dashboard, or userowsafe login. - A key that doesn't say who made it (Made before Rowsafe recorded who made keys). Make a new one.
- Apps and keys connected before agents could act as their person, until that person allows it.
They can still read and do what agents do without anyone.
What agents do without anyone, always
What never touches production, agents with write access do themselves: reading, saving Marks, backups, Proof, checks, migration previews, masked safe copies, acknowledging alerts.
What an agent can change
| Changes | |
|---|---|
| Pulse | Apply a fix, change database settings or undo a change, the index check schedule, restart the database (now or at a time), turn on backups |
| Rewind | Restore a copy, compare it with production, keep it longer or delete it, bring back rows, rewind the whole database, undo or finish a rewind |
| Updates | Install a minor update, upgrade to a new major version, undo or finish an upgrade, install the server's security updates, reboot the server, automatic minor updates |
| Standby | Promote (fail over) or remove a standby, start or stop watching a fenced old primary, automatic failover |
| Move | Schedule or make the switch, cancel, switch back, finish; restore a database whose server is lost to another server, retire it |
| Security | Turn connection pooling on or off, change security settings, fix a security finding |
| Data | Create a database for an app (Rowsafe Cloud), create a database for an existing owner, remove a database or user, turn an extension on or off, change masking rules, stop protecting a database |
| Files and alerts | Restore a database's or a server's files and undo it, change an alert rule |
| Rowsafe Cloud | Create a server, change who can connect, change a server's size, clone a database to a new server, delete a server |
The agent's describe_change tool lists them with their parameters. Creating a standby, starting a move and forking need a person (below).
Each change gets its own page under Settings → AI agents (Done by AI agents): what changed, the agent's reason, the risk, who it ran as, and the result.
Rowsafe Cloud changes
An agent can create a Rowsafe Cloud server for its app and change it later (Need a database?), telling you the price first. Good to know:
| Change | |
|---|---|
| Create a server | The page shows region, size, price per hour and the most per month (a standby doubles it), who can connect, and how it's paid. |
| Change who can connect | Opening to everyone (0.0.0.0/0 or ::/0) or a very wide range gets its own warning. |
| Change a server's size | A Mark first. A few minutes offline while the server restarts (with a standby, writes pause a few seconds). The disk never shrinks. |
| Clone to a new server | Real, unmasked data. Costs money until Rowsafe deletes it, at the time you chose. |
| Delete a server | Refused if its backup passphrase was never saved (clones aside), since its backups couldn't be restored. |
Paying for Rowsafe Cloud
- Pay as you go is on: created right away, billed by the hour (How billing works).
- Not on yet, or billed by the month (like OVHcloud Value): only owners pay. An owner's agent gives you a checkout link; the server is created once paid. An admin's or member's agent is refused: an owner creates it, or turns on pay as you go first.
Nothing is created or billed before it's paid for.
The team setting
Only an owner chooses, in Settings → AI agents (never an agent, app or API key):
| Choice | What happens |
|---|---|
| Act right away as the person who connected them (the default) | Agents act as their person. An optional monthly budget limits what agents' servers cost. |
| Within a budget | Any agent can create Rowsafe Cloud servers on behalf of the owner who chose this, and on those servers create an app database, change who can connect and clone. Within the budget. Everything else runs as the agent's own person. |
If that owner stops being an owner, the setting goes back to the default (the budget stays) and every owner gets an email.
The budget and what counts
Optional with the default, required with Within a budget. Set it in US dollars in Settings → AI agents, which shows what agents' servers cost now.
- Counts: every live Rowsafe Cloud server an agent created, clones and standbys included, at the most a month can cost; and what an agent's resize added to a person's server.
- Doesn't count: servers people created, and servers waiting for payment.
A change that would go over is refused, with the numbers (Over the $50 agent budget: ...). Rowsafe checks one agent change at a time per organization, so two agents can't both slip under it. Sizes not priced in US dollars are refused (a person makes those in the dashboard). To make room, raise the budget or delete servers.
When Rowsafe refuses
Nothing changes, and the agent tells you why, when:
- the person is a member, left, or the agent can't make changes yet;
- a first payment is needed and the person isn't an owner;
- it would go over the agents' budget;
- a server's key must be compared: creating or rebuilding a standby, moving or forking a database. A person checks it in the dashboard or the CLI, which keeps your bucket settings sealed to your own servers;
- it replaces or deletes data and there's no backup yet (the agent can run
run_backupand retry), or a Rowsafe Cloud server's backup passphrase isn't saved; - the password must be made in a person's browser: an app database asked for by an app signed in with Rowsafe, and always on MySQL, MariaDB and ClickHouse. Create it in Databases & users;
- Rowsafe can't confirm the person's role right now. It never guesses: try again in a minute. (When busy with several agent changes in your organization, it also turns some away; the agent retries.)
Safety nets that stay on
- The same path as your click, with the same checks. Restarts, updates and reboots still need root's OK on the server (Permissions).
- Marks first before risky changes (a rewind, dropping an index, a resize, an upgrade).
- No destructive change without a backup.
- Rowsafe types confirmations (a database's or server's name) for the agent, and the audit log says so.
- Backups are never deleted for an agent. A database it removes keeps its backups until an owner deletes them (Kept for you); so does a Rowsafe Cloud server it deletes.
- No secrets. Agents never set or see passwords, except an app database's password made on their own machine.
- No SQL or commands on production through Rowsafe, and no look inside your backups. Agents test on masked safe copies.
Rowsafe can't tell what you asked the agent for
Rowsafe checks the change, the role, the budget and the backups, not the conversation. An agent can misunderstand you, and an owner's or admin's agent can restart, rewind or delete a server. Give apps Act for you, and keys write access, only where you want that.
Know what agents did
- Audit log: [email protected] through their AI agent Claude.
- Dashboard: Settings → AI agents (Done by AI agents), and a banner on Home for the last 24 hours.
- Email: owners get a digest (at most an hour after the first change). Turn yours off in Settings → AI agents.
- The agent:
get_orgtells it who it acts as; each change answers with its result (MCP reference).
Apps and keys connected earlier
They show Can't make changes yet until their person allows it:
- An app: in Settings → Connected apps, click Let it act as me, or connect it again.
- An API key or CLI login: in Settings → API keys, click Let agents using it act as me, or make a new key or run
rowsafe loginagain.
Only that person can do it.
Stop an agent
- One app: Settings → Connected apps, Disconnect. Access ends right away.
- One key: Settings → API keys, revoke it.
- Cost: set a monthly budget in Settings → AI agents, and delete servers you don't need.
To let an agent only read, connect it without Act for you, or use a read-only key.
Limits
- Agents create servers only in Rowsafe Cloud, never in your own cloud account.
- The budget is a monthly rate, not a total: an hourly server costs money until it's deleted. Deleting a server frees its share.
- With Within a budget, resizing or deleting a server runs as the agent's own person.
- An app acts for the one organization picked when it connected; an API key for its own.
- The budget applies to AI apps and
rowsafe mcp. A read-write API key used directly (CLI, scripts) acts like the dashboard, with its maker's current rights.