AI Kanban

Trello-like Kanban with AI chat and MCP: users sign in, manage a board, and ask the agent to create/move cards — while Kanban domain rules and LLM orchestration live in separate FastAPI services. Live at ai-kanban-pearl.vercel.app.

Context

Many “app with AI” demos mash UI, business rules, and the LLM into one process. The goal here was a clear interview architecture story: service boundaries, user JWT on the agent (no service-role), near-zero cost, and a reliable demo path.

What the product delivers

Supabase login, seeded board (To Do / Doing / Done), card CRUD and drag-and-drop in the UI
In-app chat that creates/moves cards via tools; UI updates with SSE/streaming and refetch
Local MCP (Cursor) reuses the same tools against kanban-api with the user token
Real deploy: Next.js on Vercel, FastAPI×2 on Render, Postgres/Auth on Supabase

Engineering signals

Monorepo: apps/web (Next.js) + services/kanban-api and agent-api (Python/FastAPI) + mcp-server + shared kanban-client package
The agent never talks to the database: it only calls kanban-api while forwarding the user JWT
Alembic on API startup; Compose/Podman for local stack; CI with Python tests plus web lint and type-check
Explicit trade-offs: free-tier LLM (Gemini), Render free cold starts, local MCP in v1, in-process pub/sub for SSE

Portfolio / interview project. Public demo at ai-kanban-pearl.vercel.app. Source on GitHub (private — access on request).

#Python#FastAPI#Next.js#AI / MCP

View demo Source code