RecallFlow AI

Inspiration

I wanted to solve a familiar frustration: useful details get scattered across notes and conversations, and it’s hard to find them again when you need to act. I built RecallFlow AI to connect remembering with doing—so saved context can help answer questions and move tasks forward.

What it does

RecallFlow AI lets you save and search notes, find relevant information with source citations, and create, update, or complete tasks using natural-language requests. Notes and tasks persist locally in SQLite, and an activity panel shows the MCP tools used for each action.

How I built it

RecallFlow combines a FastAPI backend, SQLite storage, and a React + TypeScript browser app. The backend plans supported requests; the browser uses the official MCP JavaScript SDK to discover and call tools on an official MCP Python SDK server over Streamable HTTP. The server exposes tools for saving and searching notes, managing tasks, and retrieving a productivity summary. Answers are grounded in retrieved data and can cite the source note. The optional language-model adapter runs server-side; the app also works in deterministic demo mode.

Challenges I ran into

The trickiest challenge was making browser-based MCP sessions reliable. FastAPI’s route mounting initially caused redirects, and browser CORS hid the MCP session ID needed for later tool calls. We fixed the endpoint routing and exposed the required headers. We also designed the assistant to avoid inventing answers or acting on ambiguous task matches.

Accomplishments that I'm proud of

1. Built a working end-to-end browser assistant with persistent notes and tasks stored locally in SQLite.
2. Integrated eight MCP tools over Streamable HTTP, with real browser-side tool discovery and calls.
3. Grounded answers in saved notes, showed source citations, and avoided guessing when context was missing or task matching was ambiguous.
4. Made tool activity visible so users can inspect actions and results.
5. Verified the project with 18 passing tests and a successful production build.

What I learned

Building RecallFlow taught me that a reliable MCP integration takes more than defining tools: the browser and server must handle initialization, tool discovery, session IDs, and CORS correctly. I also learned to ground answers in saved data, show which tools ran, and avoid guessing when information is missing or ambiguous.

What's next for RecallFlow AI

Next, I’d like to make RecallFlow more useful in everyday life by adding semantic search and opt-in integrations with calendars and other productivity tools. I’d also explore secure cloud sync and a real Alexa+ integration if access becomes available, while keeping users in control of their data and actions.

Built With

• context
• fastapi
• model
• protocol
• python
• react
• sqlite
• typescript
• vite