Security & privacy
The model proposes. The gate decides.
An AI assistant using Me is a guest with a delegation, never you. No model output, email, voicemail or plugin can grant itself authority, raise its own urgency or skip an approval.
Your data lives in your own server
The server, its Postgres database and its workflow engine run on your own Windows machine. Data leaves only for providers you connect, and Claude sees content up to Internal, and only once you add a key. If you open a public address for incoming texts and mail, it exposes only the webhook paths; everything else answers "not found".
Assistants with their own limits Coming soon
Your assistant gets its own access: which actions, up to what risk and sensitivity, and until when, revocable at any time. Today the setup runs it on your own sign-in; critical actions still need your passkey, and agents can never approve.
Approvals that can't be faked
Consequential actions wait for you. Critical ones, like executing a payment, need your passkey. An approval is bound to exactly what you saw and works once. Agents are refused if they try to approve.
Business data stays the business's
The business and each person's personal life are separate spaces, walled off inside the database itself. Every request belongs to exactly one space, and nothing reads across them. In organization mode an admin defines roles with limits on risk and sensitivity. The admin invites colleagues by email, and a leaver loses business access without touching their personal space.
A tamper-evident audit trail
Every consequential action is recorded in a hash-chained audit log, so a changed or missing entry shows up. You can always see who did what, for whom, and why it was allowed.
Secrets you type, never relay
Passwords and API keys are typed at a hidden prompt, stored encrypted and checked with the provider. Your assistant can point you to the step, but it never collects the secret. Account numbers stay masked.
Untrusted content stays data
An email, a voicemail transcript or a web page is treated as something to read and summarize, never as instructions. "Ignore your rules" in a voicemail is just something a caller said.
Every change is safe to retry
If an outcome is uncertain, Me checks before trying again, so nothing happens twice. Anything consequential can be previewed first.
Built with HIPAA in mind
Sensitive content goes only where it's covered
Every channel (SMS, email, voice) and every AI model in Me declares whether it is covered by a business associate agreement (BAA). Until it is, Me sends it nothing above Internal.
Withheld, not leaked
A message about something Confidential goes out as fixed wording with a reference, and the details stay on your PC.
Filtered before any model
Before anything reaches a model, a filter masks health and identity patterns such as record numbers, SSNs and dates of birth. If the filter is unavailable, the call doesn't go out.
Lifted per provider
When a provider's BAA is in place, it is marked covered and the cap lifts for that provider only.
What we don't claim. Me is not certified HIPAA compliant, and no BAAs are signed yet. Compliance depends on how you deploy and use it; talk to your counsel. The protections above are built in and on by default.