Why everyone in Baseboard starts read-only
2 min read
The first support hire is usually when an admin panel's permissions get tested. You want them to look people up, see what they pay and check whether the last deploy broke sign-in. You don't want them issuing refunds on day one, and you'd rather nobody could drop a table by accident.
Most homemade panels have one role: if you can sign in, you can do everything. This post is about the rule Baseboard uses instead, and the one part of it people find odd at first.
Reading and changing are separate
Every action in Baseboard has a level for each member, Allowed or Not allowed. The actions fall into three groups:
| Group | Examples | New members start |
|---|---|---|
| Dashboard | Managing connections, creating projects, alert rules | Allowed |
| Reading data | Users, invoices, rows, files, deployments, errors | Allowed |
| Changing data | Refunds, bans, cancellations, rollbacks, cell edits, write SQL, file deletes | Not allowed |
So a new member can open every page in their projects and change nothing. To let someone refund, an owner opens their permissions and sets Refund invoice to Allowed. That's it. There's no role to design and no policy language to learn.
Owners don't get a pass
This is the part that surprises people. Owners start with the same levels as everyone else, so the person who made the workspace can't refund an invoice until they allow it for themselves.
We did this on purpose. A permission system with an "owner can do anything" escape hatch means the most-used account in the workspace is also the one with no guard rails. Making owners tick the box for themselves turns every risky change into a decision somebody made, on the record. Owners can always change their own levels, so nobody gets locked out.
Agents get the same rules
An AI agent connected over MCP, or the assistant built into Baseboard, acts as a member. A token can be narrowed further to some tools and projects, but it can never do more than the member who made it. If you can't issue refunds, neither can your agent. We wrote more about that in Letting an AI agent near production.
Everything is written down
Every read and every change lands in the project's audit log with who did it: a member, the assistant acting for them, or a token. When someone asks why a subscription was cancelled, the answer is a filter rather than a Slack thread.
The full list of actions and levels is in Permissions.