Skip to main content

How Meggy keeps your books safe

The short version: it runs where you decide, every write is made as a named member and recorded, and the database itself refuses what a role may not do. The longer version follows.

Version of 5 November 2026.

Where it runs

The server edition is something you host: a PostgreSQL database you own, an API and a static site in front of it. The reference deployment uses Firebase Hosting, Cloud Run and Cloud SQL. The database is plain PostgreSQL; sign-in, document storage off the server's own disk and the queue behind Telegram and email intake use Google's services today.

The desktop edition embeds the database in the application on your own computer. Its network guard refuses every outbound connection to a host you have not allowed by name — including its own update check.

A server install is backed up the way its database is, which is the operator's to set up; the operator can also export an active workspace as an archive stamped with the release it came from. The desktop edition copies its data directory before every update and daily, and keeps the copies beside it.

Who you are

On the server edition you sign in with Firebase Auth, by email or Google, and every request carries a token the API verifies against the deployment's own project. On the desktop edition the shell mints a token per launch and hands it to the embedded API itself; the interface never holds a credential.

A workspace's members are invited by its owner from the Members screen, with a one-use link; nobody joins a workspace by signing up. Whether new accounts can be created at all is a setting of your Firebase project. Transferring ownership or changing a workspace's lifecycle additionally requires a sign-in within the last five minutes, and a token refresh does not count.

What the database enforces

Every transaction opens by binding the caller's membership and the capability it is exercising, and the row policies deny by default: a read with no bound context is refused, not admitted.

Access is per book — view, edit or administer — and per subsystem, granted by the owner and checked in SQL routines the application cannot bypass. A dashboard's figures come from database reads too, under the same book, subsystem and row visibility as every other read, so a member's totals count only the books, accounts and areas they hold, and say so when something was left out. An auditor's membership is structurally excluded from every writing lane.

A posted entry is immutable. Its postings balance to zero in the book's currency, a posting against an account carries that account's own currency beside the book's, and a correction is a reversal that leaves the original in place.

What is on record

Row triggers capture every change into an audit log the application's own credential can only append to. Reading it is a separate lane with a separate database login, so the runtime credential cannot read the log as the auditor, and the auditor cannot write.

A member can be invited as audit-only: they read the log and nothing else, and no ordinary screen can be opened for them at all.

Secrets

AI provider keys and bank credentials are sealed with a versioned keyring into a separate secrets table, never as a column beside the thing they belong to. A runtime with no keyring refuses to start rather than run unsealed.

A saved key never comes back to the screen; the page says only whether one is stored. No provider key is read from the environment; the keyring that seals them is mounted as a file rather than set as a variable, because a variable is inherited by every process the service starts.

AI

You bring a provider's key of your own, or point Meggy at a local model, which costs nothing and is exempt from the spend limit only when it is genuinely local. On Personal or Pro the AI the plan includes is on by default, which runs on the service's own key and asks the provider for endpoints that keep no data. Every call is admitted against the spend limit and recorded before the provider is contacted, so a limit can stop it.

A model may propose the fields of an entry. It may not choose an account it has not been shown, and it may not write the ledger; you post, or you do not, unless the workspace owner has let the chat record a fully trusted cash receipt, up to a limit the owner sets. A single server-side switch turns every AI feature off at once.

Documents

Uploaded documents — statements, receipts and the rest — and account images live in two separate stores, and account images may never fall back to the documents' store; both are served only behind sign-in and membership, by an unguessable id, never at a public address.

Reporting a problem

If you find a security problem, write to the contact address at the foot of this page, with what you did and what you saw. Please test only your own account, do not read other people's data, and give us a reasonable time to fix it before saying anything publicly. We will not pursue anyone who acts in good faith along these lines. There is no paid bounty.