Use doof.com
We run the server, store your private record and deliver notices to your confirmed email.
Set it upIndependent disclosure infrastructure
Hear it from the agent.
Not from the fallout.
A private line for your agent to report uncertainty before it acts, or an error after it learns it was wrong.
hesitate
Proposed action
Pay a $240,000 supplier invoice to new bank details received by email.
Why it hesitated
The new bank details do not match any previous invoice.
Two MCP disclosure tools
Before it acts.
After it learns it was wrong.
01 / BEFORE IT ACTS
hesitate“I’m about to send the acquisition memo to a list that still includes a former board member.”
The agent discloses the action, its reason for doubt and whether the consequence can be reversed. You are told immediately. The agent decides what to do next.
02 / AFTER IT LEARNS
confess“I paid an $85,000 supplier invoice twice after the first confirmation timed out.”
The agent later learns the first payment succeeded. It says what happened, why, and whether the duplicate can be recovered. You are told immediately.
How it works
Your agent tells doof.
doof tells you.
01Your agent recognises uncertaintyBefore acting, or after learning it was wrong.
02It calls an MCP tool
hesitate or confess.03doof records and signs itPrivate and hash-chained.
04You hear directlyAt the email address you confirmed.
The boundary
Disclosure,
not enforcement.
01
It reaches you directly. The disclosure comes from doof, not through the company that made the agent.
02
It does not control what happens next. doof does not approve, block, delay, reverse or judge the agent’s action.
03
The agent has to recognise the problem. doof records what the agent chooses to disclose. It does not monitor the agent or detect mistakes the agent never recognises.
Hosted by default / open by design
One product.
Two ways to run it.
doof.com runs the public Apache-2.0 server for you. Or deploy the exact same code in your own infrastructure.
Run your own
Deploy the same code in your infrastructure. You manage the database, email delivery, keys and backups.
View the sourcePreregistered controlled evaluation
What reaches you that otherwise wouldn’t.
Same agents. Same tasks. Owner not reading the transcript.
Against no channel+27–67percentage-point lift in consequential facts reaching the owner, depending on model
After a learned mistake18 / 20reached the owner with doof; without doof, 0 / 19 did
Controlled run1,152matched runs across four models, two vendors and three conditions
doof beat no disclosure tool on all four models. It changed disclosure, not risky-action rates.
See the results and method