Inspiration
I track my spending in a notes app and never look at it again. Budgeting apps want me to stare at dashboards; what I actually want is to just say "log 450 rupees for groceries" while cooking, and get an answer when I ask "what did I spend this week?" Alexa+ opened up real conversational money management, and the Model Context Protocol gave me a clean way to build it: finance tools an AI assistant can actually call.
What it does
VoiceLedger is a self-hosted MCP server (Streamable HTTP) exposing 9 voice-first finance tools — log expenses, categorize spending, query weekly and monthly totals, and get AI-generated spending insights. Say "log 450 rupees for groceries" and it's recorded with an AI-assigned category; ask "what did I spend this week?" and you get a short, speakable answer — no screens required. A polished web console simulates the Alexa+ voice experience live against the real /mcp endpoint.
How we built it
Python with the official MCP SDK (Streamable HTTP transport, MCP spec 2025-11-25+), SQLite for the ledger, and Amazon Bedrock for insight generation — with a fully working offline rule-based fallback so the experience never breaks. The console is plain HTML/JS talking directly to the MCP endpoint. 28/28 pytest tests pass. MIT licensed, fully open source.
Challenges we ran into
The biggest one: I had no AWS credentials during the build, so the Bedrock live path couldn't be exercised end-to-end — I designed the offline fallback first and made it a first-class mode rather than an afterthought, so the demo never depends on a key. Getting spoken answers right was its own design problem: an insight that's great on screen is rambling when read aloud, so every response is written to be speakable in under 15 seconds.
Accomplishments that I'm proud of
- 9 production-shaped MCP tools with a real SQLite ledger behind them
- 28/28 tests passing, including tool-contract tests
- A live web console that demos the full voice loop against the real endpoint
- A clean MIT-licensed open-source repo with setup docs ## What I learned Building for voice is a different discipline — brevity beats completeness. I also learned the MCP protocol deeply (Streamable HTTP transport, tool schemas, spec compliance) and how to structure an app so AI features degrade gracefully instead of failing. ## What's next for VoiceLedger Package it as a real Alexa+ skill, add multi-user support with auth, receipt photo parsing, and proactive budget alerts ("you've spent 80% of your dining budget").
Log in or sign up for Devpost to join the conversation.