Alderfi

Status: Pre-alpha. Vision, architecture, and a minimal MCP scaffold are published here. Active development begins after Help Tech’s May 1, 2026 go/no-go decision. First real commits (if approved) land the week of May 4.

Alderfi is the first AI-native, open-source personal finance platform. When OpenAI acquired Hiro Finance in April 2026, thousands of privacy-conscious users lost their “AI CFO” with four weeks of notice. This project is the open, self-hosted answer: a conversational financial assistant where AI is the core interface — not an afterthought — built on the Model Context Protocol so it works with Claude, Cursor, your own agents, or our web UI as equal clients.

Philosophy

• AI-native, not AI-bolted-on. Every capability is an MCP tool. Conversation comes first; forms come second (and sometimes not at all).
• Open source, forever. Apache-2.0. We will not relicense the core. If we ever do, you can fork the day before.
• Self-hosted and local-first. SQLite + your hardware by default. Cloud is an option, never a requirement.
• Layered AI. Merchant rules handle the easy 70–80% with zero latency. A local LLM (Llama 3 8B via llama.cpp) handles the ambiguous 15–25%. Cloud LLMs are opt-in for edge cases. Your data only leaves the box if you say so.
• Read-only banking by default. SimpleFIN-first (proven at scale with Actual Budget). Plaid as a premium option for hosted users. Your banking credentials never touch our servers.

What’s in this repo today

Just the scaffold. One MCP tool (list_accounts) returns mock data. This proves the protocol surface is real — no banking connectors, no AI categorization, no persistence yet. Those arrive in Phases 1–3.

Quick start

npm install
npm run dev

Then point an MCP client at the server (stdio transport):

npx @modelcontextprotocol/inspector npm run dev

You should see one tool, list_accounts, returning three mock accounts.

Roadmap

High-level phases (detailed architecture in ROADMAP.md):

• Phase 1 (weeks 1–2): MCP server + SimpleFIN/CSV connectors, transaction sync, SQLite store.
• Phase 2 (weeks 3–4): Layered categorization (rules → local LLM → cloud opt-in), anomaly detection, spending insights.
• Phase 3 (weeks 5–6): Conversational UI — natural-language budgeting, proactive alerts, chat-first queries.
• Phase 4 (week 7): Public alpha, Show HN / Product Hunt launch.

Development begins in earnest after Help Tech’s May 1, 2026 go/no-go decision. First commits land the week of May 4 if approved.

License & why Apache-2.0

License: Apache-2.0. Why Apache-2.0 over MIT for this project specifically: