Why you can trust it
Trust you can hand to your boss.
It learns only what you teach it, asks before it acts, and keeps every step on the record. Your data stays inside AWS and never trains public models.
Runs inside AWS
Everything on the record
Never trains public models
Customer bank details
VAULT
Locked away, kept separate, and left out of every share. No exceptions.
AUDIT_LOG · append-onlylive
Needs your OK A person signs off
Your agent wants to email 1 report to your outside auditor. Nothing leaves until you say yes.
What it learns
It learns what you teach it. Nothing else.
An agent that does your work has to learn how you work. You decide what that includes, and you can take it with you or make it forget.
It learns from
- Tasks you record and save. You press Record, you review the steps, and your agent follows a skill only once you save it.
- Your chats with it. It remembers what you tell it, so you don't explain the same thing twice.
- The apps you connect. It reads them when it's doing a task for you, with your access and nothing more.
It never
- Watches your screen when you aren't recording. Recording starts and stops with you, and screenshots are deleted once they've been read.
- Shares what it knows about you without your OK. Your memory is private by default, and secrets in the vault are never shared at all.
- Uses your data to train public models. Inference runs on AWS Bedrock, which doesn't train on it.
Export everything it knows about you, or have it forget everything, any time from Settings.
0
of your data trains public models
0%
of changes on the record
0
messages sent for you without your OK
Carve Protect
Simple promises you can count on.
Private by default
- Stays in your workspace. No one else can see it.
- Never used to train public models
- Secrets live in a vault that's never shared
A person approves
- Every send, update and new record waits for a click
- Nothing is shared with a teammate until you approve it
- If a send can't be confirmed, you decide. It's never retried on its own.
Everything on the record
- Every read, change and approval is logged, with who and when
- Append-only by design: nothing in the app edits or deletes a record
- Memories are set aside, never silently erased
Models & infrastructure
The guardrails live in CarveAI, not in a model.
Your memory, your skills and every rule above sit in our layer, so they hold whichever model does the work.
Inference stays inside AWS
- Models run on AWS Bedrock, so your company's data stays inside AWS
- Your data is never used to train public models
You control the model pool
- Models are added deliberately, never auto-adopted
- Set which are allowed, per agent or company-wide
Every choice on the record
- Which model ran each task is written to the audit log
- Switching models never changes what your agent knows
On the record
See exactly what happened, every time.
Every read, change and approval is written down, and nothing can be edited away after the fact.
audit_events · append-onlylive
10:14:02APP_READ · outlook#9241
10:14:03SKILL_FOLLOWED#9242
10:14:07CARD_SHOWN · sheets#9243
10:14:12HUMAN_APPROVAL#9244
10:14:13WRITE_CONFIRMED · sheets#9245
10:14:21HUMAN_REJECTION#9246
Illustrative stream: sample audit events.
Who we are
A small team building AI you can actually trust
Questions we get asked
Plain answers, no hedging
- LearningWhat does it learn about me?
- What you show it and what you tell it: the tasks you record and save, and what you say in your chats with it. It doesn't watch your screen when you aren't recording. You can export everything it knows about you, or have it forget everything, from Settings.
- ControlCan it send or change things without me?
- No. Every email, update and new record waits on an approval card that shows exactly what will happen. If it can't confirm something went through, it tells you instead of trying again.
- TrainingDo you use our data to train AI?
- Never for public models. Inference runs on AWS Bedrock, which doesn't train on your data. Down the road, your data may fine-tune a model that's private to your company, on your data, for you. It's never used to train anyone else's model.
- Your dataIs our memory really ours?
- Yes. It stays in your workspace, it's private to each person by default, and nothing is shared until a person approves it.
- ModelsWhich AI models do you use?
- As many as fit the job. CarveAI routes across 30+ frontier and open-source models on AWS Bedrock, and you set which are allowed. Switching models never loses what your agent knows.
Want the details?
Our full privacy and security write-up is available on request.
Request the full write-up →An agent you can actually trust with real work.
A 30-minute walkthrough. See it for yourself.