Built so that a mistake by the AI cannot become a sent email

The approval gate

The language model can never approve anything. Sending an email and cancelling a meeting always need your tap on a card that shows the exact message and recipients. Approval goes through a single-use database latch with a 2-hour expiry, so a card cannot be reused. New and moved meetings can be set to auto-approve in Settings, only when the slot has no conflict.

• Gated tools tell the model "NOT happened yet". Typing "send it" cannot trick it into claiming something was sent.
• Success is reported only after Google's API confirms it.
• Sends from chat wait 10 seconds and can be undone.
• Recipients of drafts and nudges are chosen by code, never by the model.

Prompt injection

Email content, attachments and tool results are marked as untrusted data in every prompt. Memories can only be saved when the text appears word for word in what you typed, so an email cannot plant one. The brief's AI gets no tools and any id it invents is dropped. Even if a model is fooled, every outbound side effect still needs a human click.

Permissions, and why

Momo uses Google sign-in and never sees your password.

• gmail.readonly: search, read and summarise mail you ask about, build the brief and triage.
• gmail.send: send an email, only after you tap Send.
• gmail.compose: save drafts to Gmail Drafts.
• gmail.modify: archive mail you approve. Never deletes.
• calendar.readonly and calendar.events: list events, find free time, create, move or cancel after approval.
• contacts.readonly and contacts.other.readonly: work out who "Rahul" is.

What is stored and for how long

Your Google refresh token is encrypted with AES-256-GCM and used only server-side. Conversations are encrypted at rest and purged 30 days after the last message. Triage items keep sender, subject and a snippet, encrypted, for 14 days. Promise text and memory facts are encrypted with an HMAC fingerprint for de-duplication. Your writing-style profile stores style, not email text. The audit log keeps 180 days and holds no email bodies.

AI processing

Momo uses Azure OpenAI models under Microsoft's enterprise terms, which do not use customer data to train models. Momo does not train its own models on your email. Attachments are parsed in memory with size and entry caps, never written to disk, and no macros or formulas are executed.

Where it runs

Application servers run on Google Cloud in Mumbai (asia-south1). The database runs on Azure in Central India with TLS required. AI model calls go to Azure OpenAI in the United States (East US 2) under Microsoft's enterprise terms. Static pages are served by Cloudflare.

Leaving

Settings has Disconnect Google, which revokes Momo's access at Google, and Delete everything, which removes conversations, memory, drafts records and stored data. Google verification of the app is in progress; until it completes, Google shows an "unverified app" screen at sign-in.