Bellion: a private AI finance agent
Everyone talks about agentic AI; fewer people show it doing a real job. Bellion is my attempt: an assistant that watches my own finances, briefs me every morning and answers questions over Telegram, with the model and my financial history kept on my own machine.
- Language model
- Local
- Memory
- AES-256
- Daily briefing
- 7:00 am
- Interface
- Telegram
The problem
Budgeting apps show you charts after the fact. I wanted something closer to a sharp analyst who checks my accounts every morning, flags anything odd, remembers my goals, and can answer a question like "how did my Uber spending compare to February?" in plain language. I also didn't want my bank history sitting in someone else's cloud.
How it works
- The agent: OpenClaw runs a Gemma model locally and decides which tools to use.
- Data: Plaid provides read-only bank and card transactions, stored in a local SQLite database the agent can query with SQL.
- Memory: a local vector store lets it recall past conversations and goals ("save an extra $500 next month for a trip").
- Interface: Telegram, restricted to my account only. A scheduled job sends a formatted morning briefing covering debt progress, savings goals, yesterday's spending and a "Daily Zero" check.
- Alerts: small background scripts message me about transactions over $500, or if the laptop running the agent is about to die.
What runs where
| Part | Runs on |
|---|---|
| Language model and agent | My machine |
| Chat history and memory (encrypted) | My machine |
| Bank data | Fetched from Plaid, stored locally |
| Messages | Telegram |
| Web search, image model | Cloud, only when used |
"Private" here means my records and history live on my machine, encrypted. It doesn't mean nothing ever touches the internet, and I think it's worth being precise about that.
Current status. The agent and its Telegram channel work today. The separate encrypted gateway stores every message correctly but still sends a placeholder reply; wiring the agent through it, so every message is encrypted end to end on my side, is the next step.
What I'd do next
- Route the agent through the encrypted gateway, so there's one audited path in and out.
- Add a human-approval step for any action that changes money or accounts, enforced in code rather than in the prompt.
- Write tests for the gateway covering signature checks, encryption and the alert scripts.