Pennyworth

The app is zero-knowledge at rest. On sign-up the phone generates a 256-bit AES-GCM data key and wraps it twice, once under a PBKDF2-SHA256 key derived from the password at 600,000 iterations, and once under a printed recovery code. Only those two wrapped blobs are stored.

Amounts, account names, merchants and notes are encrypted field by field on the device before they leave it, so the amount column in Postgres holds ciphertext rather than numbers. Balances are never stored at all: the phone recomputes them from an opening balance plus every entry. There is no application server between the app and the database, and Supabase row level security is the entire authorisation layer.

The AI features are not zero-knowledge, and it would be dishonest to imply otherwise. The device decrypts, then posts plaintext to the app's own API, which forwards it to Anthropic. Asking the assistant a question sends a summary built on the phone: net worth, account names and balances, six months of income and spending, top categories, budgets, goals, recurring payments and investment positions. Importing a statement uploads the whole PDF. Nothing is persisted server-side, as those routes contain no database writes and no logging, but the claim that the server can never read your figures stops being true the moment an AI feature is used.

There are five API routes, all POST, all gated behind a Supabase JWT and a subscription check. Claude Haiku 4.5 handles four cheap structured jobs: parsing a sentence into an entry, classifying intent, turning a search phrase into a filter the device then runs locally, and mapping and categorising statement rows. Each uses forced strict tool use, so the model returns a validated object rather than prose. Claude Sonnet 5 handles only the conversational assistant, with the financial summary marked as an ephemeral cache block so it is not re-billed every turn.

Two Vercel cron jobs run once daily and neither touches a language model. One posts due recurring payments, the other snapshots prices and warms foreign exchange rates. Everything else runs on demand on the device.

Three services are live and keyless in production: CoinGecko for crypto prices, Frankfurter for foreign exchange, and the Anthropic API. Etherscan, Alchemy, Helius and Finnhub are built but key-gated and issuing no requests. Coinbase, Kraken, Gemini, Binance and Trading 212 connect on each user's own read-only key, so they cost nothing at any number of users. TrueLayer open banking is a working scaffold and inert until configured, with real work still to do before it can ship.

Nothing today, because nothing is launched. Everything sits on a free tier and the Anthropic API has cost $0 to date.

At launch it changes sharply, mostly because of open banking.

That splits into about $53 of fixed infrastructure, $406 of open banking and $100 budgeted for AI. TrueLayer alone is 73% of it and does not move with user count. The $405.75 is £300 a month converted at 1.3525.

At £6.99 a month per subscriber, infrastructure breaks even at roughly 60 subscribers on gross revenue, or roughly 70 after Apple's 15% Small Business Program commission. Without TrueLayer, break-even drops to about 8.

Two figures in that projection are estimates rather than measurements. The cost of PDF statement imports is unmeasured, and Anthropic bills a PDF as extracted text plus one page image per page, so it scales with page count. The $100 for AI is a budget rather than a measured bill.

Claude and Codex are subscriptions used across everything, including this blog, so they are listed as part of the stack rather than charged to this project. The same goes for the Notion workspace.