Security and approvals
It holds your keys, so it is built to be careful.
Fumu treats every new connection and web page with care, and limits what each part can reach. Reading is free; changing things is earned.
The policy gate
Run it, ask you, or refuse.
Before every action, Fumu checks how risky it is, how much trust you have given that capability, and the rules you have written. Trust is raised one capability at a time, as each one proves itself.
| Level | Fumu may | For example |
|---|---|---|
| 0Observe | Read and summarise only. | "What's in my inbox?" |
| 1Suggest | Draft, then ask before changing anything. | Drafts a reply and waits for your approval. |
| 2Act with notice | Make changes that can be undone, tell you afterwards, and allow undo for a while. | Archives newsletters, accepts a meeting. |
| 3Autonomous within budget | Act without asking, inside limits you set. | Reorders groceries under €40. |
Rules that never relax, whatever the level
Approvals
How an approval works.
Fumu asks
You get a card with what it wants to do, why, and exactly what will change: the code, the email text, the amount.
Fumu waits
The task waits for up to 24 hours by default, reminds you once, then cancels safely. Nothing is lost while it waits.
You decide
Approve, Edit or Reject. Edit lets you change the details, the wording of an email say, before it goes ahead.
It is written down
Every decision is recorded. Approve the same thing often and Fumu suggests a rule for it.
Safeguards
Least privilege, everywhere.
Only what each job needs
Every capability says up front what it needs to read and change. Anything else is refused.
Your passwords stay hidden
Keys and passwords are kept locked away and used only at the moment of an action. The AI never sees them.
Outside text is treated with care
Words inside an email or a web page can never trigger a risky action on their own, and plans that rely on them get a second check.
Kept apart
Each capability runs on its own, with no access to the others or to the internet beyond what it needs.
Approved additions only
New capabilities come from a trusted list or after your approval, and never update themselves silently.
A record you can search
Every action is recorded: who asked, which agent, the result and the approval. Nothing is ever edited out.
A kill switch
"Fumu, stop" cancels every running task and drops every agent to Observe.
Only you get in
Only your account can talk to Fumu, with a PIN for high-risk confirmations, and the console is protected by a passkey.
Questions about security?
Tell us what you would need to see before trusting an agent with your inbox or your repositories.
Contact us