Agentic Calendar

The Problem & Target User
Modern calendar apps — Google Calendar, Outlook, Apple Calendar — are powerful but fundamentally passive tools. They store what you tell them, but they do nothing to help you think about your time. For students balancing coursework, projects, and extracurriculars, and for professionals juggling meetings, deadlines, and personal commitments, the real bottleneck isn't a lack of calendar features — it's the cognitive overhead of deciding what goes where, when, and why. Our target users are time-pressured individuals who already live inside a calendar but spend far too much mental energy maintaining it manually. How AI Is Used Agentic Calendar replaces manual scheduling with a conversational AI agent powered by Claude. The AI is not a chatbot bolted onto the side of a calendar — it is the primary interface. It reads your existing events, understands your schedule context and personal preferences learned over time, and takes real actions: creating, moving, and deleting events via natural language. For more complex planning, a Workspace mode lets the AI propose entire batches of draft events — shown at reduced opacity on the calendar — which users can review and accept or reject before committing. The AI also manages a live todo list, resolves conflicts automatically, and personalizes its suggestions through a persistent memory of your habits and work patterns stored across sessions. What Inspired Us The inspiration came from frustration with the gap between AI assistants that talk about your schedule and tools that actually manage it. We wanted to build what calendar apps should have always been: a system where you describe your week in natural language and the AI handles the rest. The emergence of reliable tool-calling in large language models made this genuinely achievable for the first time.

How We Built It We built Agentic Calendar over five days using Next.js 16 as the full-stack framework, Supabase for authentication and a PostgreSQL database with row-level security, and FullCalendar v6 for the calendar rendering engine. The AI layer is built on the Anthropic Claude API with a skill-based tool system — each capability (create event, move event, propose draft batch, manage todos, update user context) is a discrete tool with injected instructions that guide Claude's reasoning step by step. We integrated Google Calendar, Microsoft Outlook, and Apple iCloud (via CalDAV) as external sync sources, and added cross-user calendar sharing with a Supabase RLS policy architecture that avoids any data duplication. The UI uses shadcn/ui and Tailwind CSS 4, with TanStack Query managing all server state and optimistic updates so the calendar feels instant. Challenges We Faced The hardest challenge was making the AI behave reliably as an agent rather than a chatbot. Early versions would describe what it was about to do instead of actually calling tools, or would call propose_workspace_draft once per event instead of batching them all into one call — breaking the entire workspace panel. We solved this through carefully engineered system prompt instructions with explicit CRITICAL rules. Recurring event deletion was another subtle bug: the AI would loop through every occurrence individually rather than deleting the series by its root ID, which we fixed by surfacing is_recurring and recurrence_rule fields in search results so Claude could reason correctly. On the infrastructure side, unifying three external calendar integrations (Google OAuth2, Microsoft Graph OAuth2, Apple CalDAV) into a single consistent architecture required significant refactoring — each platform has its own auth model, token refresh flow, and event format. Finally, building live calendar updates that feel optimistic (the calendar updates before the server confirms) while staying consistent required careful coordination between TanStack Query cache mutations and background invalidation.