Permissions
Control that is enforced, not advised
Every one of the 378 actions is set independently to always allow, ask every time, or disabled. The defaults are deliberate, and the strictest setting is stronger than a refusal.
The problem
Most AI governance is a document
A policy that says what an assistant should not do is only as good as the assistant's willingness to comply with it. That is a strange thing to rely on when the action is writing to your CRM.
Read a threadslack.read_threadAlways allow
Search the CRMhubspot.query_recordsAlways allow
Draft an emailgmail.create_draftAsk every time
Create an issuelinear.create_issueAsk every time
Run repo commandssandbox.git_execDisabled
Writes always stop for a human. There is no auto-approve mode anywhere in the product, on any plan.
What it does
Always allow
The default for a plain read. It runs without stopping to ask, because reading was never the part that needed permission.
Ask every time
The default for all 141 writes and the 11 reads flagged sensitive. The run halts and shows you exactly what it wants to do before it happens.
Disabled
Not a refusal. The action is never offered to the model at all, so there is nothing to talk it out of.
The guarantee, stated plainly
There is no auto-approve mode anywhere in the product, on any plan.
- Write tools are off entirely until an admin turns them on
- Even then, each write stops for a human before it runs
- Unattended work is gated separately and more strictly than chat
- Every permission change is written to the audit log
The rest of it
Give your AI a company to work for
Connect your tools, and every agent you already use starts the day knowing how your company actually works.