Desktop app · macOS and Windows · free
Slow and steady.
SlothSpend reads a year of your bank statement and finds the subscriptions you forgot, the prices that crept up and the charges that came twice — with open models running on your own computer.
No account. No cloud. Your statement never leaves your computer.
How it works
Audit, ask, remind, verify.
Each step is something you can see and check. Every number says which model produced it, and every answer shows the tools it used.
-
Audit
Drop in a CSV export. Gemma tidies the merchant names; TabPFN learns which charges recur and forecasts the next 30 days with a range, not a guess. Double charges, price creep and overlapping services get flagged.
-
Ask
“What if I cancel StreamFlix and WatchNest?” Ask Sloth answers with typed tools — including a live TabPFN forecast — and shows each call it made. It drafts refund and cancellation emails from your own charge history.
-
Remind
Pick a subscription and SlothSpend reminds you a few days before it renews — a desktop notification while the app is open, or the next time you open it. Keep it, snooze it, or cancel it yourself.
-
Verify
Said you cancelled? Import your next statement and SlothSpend checks the charge really stopped. If it didn’t, a refund request is ready to send.
Inform, never act.
- It never cancels a subscription or logs in anywhere for you.
- It never contacts a merchant or sends an email. It drafts; you send.
- It never invents a link or a phone number — drafts are checked before you see them.
- It never says it did something it didn’t — not even when the model wants to.
Why open models
Your bank data stays yours.
A year of transactions says where you live, what you eat and who you love. That only belongs on your own computer — and open models are what make a useful assistant possible there.
Gemma names, explains and drafts
Any Gemma model you have in Ollama. It cleans “ACH DEBIT IRONTRAP FITNESS MEMB 7731” into “IronTrap Fitness”, answers questions with tools, and writes notes and drafts — and code checks every number it writes against your data.
TabPFN decides what recurs
A tabular foundation model that learns from a handful of examples in context. Mark one charge “not a subscription” and it re-scores in seconds. Add a free Prior Labs key to upgrade to TabPFN-3.5.
Privacy
Where your data goes.
| Data | Goes to | When | Default |
|---|---|---|---|
| Transactions, merchants, amounts, chats, reminders | This computer Local database in the app-data folder | Always | — |
| Merchant descriptions → Gemma | This computer Ollama on 127.0.0.1 | Import | On |
| Reminder and email facts → Gemma | This computer Ollama on 127.0.0.1 | A reminder fires or you ask for a draft | On |
| TabPFN inputs | This computer Local PyTorch | Import and chat | On |
| TabPFN v2 weights | Internet · one time Hugging Face | First run | Automatic |
| TabPFN-3.5 weights | Internet · one time Prior Labs (license check) + Hugging Face | Only after you add a key | Off |
| Gemma weights | Internet · one time Ollama registry | First run | You start it |
| Span timings, token counts, tool names — no prompt or answer text | Internet · opt-in Sentry | Only if traces are on | Off |
After the one-time model downloads, SlothSpend makes no network calls at all unless traces are on. A test in the repository blocks every outbound connection while the whole app runs, to keep it that way.
Download
Get started in a few minutes.
macOS (Apple Silicon)
Download .dmg
This hackathon build isn’t notarized. Drag SlothSpend to Applications, then
right-click it and choose Open — or run
xattr -cr /Applications/SlothSpend.app once.
Windows (x64)
Download .exeThe installer isn’t code-signed, so SmartScreen may warn you. Choose More info → Run anyway.
First run
- Install Ollama and open it.
- Any Gemma model works. If you have none, SlothSpend offers to download one.
- TabPFN v2 downloads automatically, once (about 70 MB, no account).
- Optional: paste a free Prior Labs key to upgrade to TabPFN-3.5 — a model licence, not an app login.