Schedule AI-powered tasks that run automatically on a recurring basis. Each task executes a prompt using Claude, optionally enriched with live web search results, and stores the output for later review.

Started building this because the native task/schedules UX in Grok and ChatGPT wasn't very good.

• Recurring schedules — hourly, daily, weekly, monthly, or one-time execution
• Web search context — optional Brave Search integration injects current information into prompts
• Structured output — each run produces a title, summary, and full markdown body
• Activity feed — filterable dashboard showing all runs across tasks
• Token tracking — input/output token usage recorded per run
• Local-first — all data stored client-side in IndexedDB via Dexie

git clone https://github.com/razgraf/cron.git
cd cron
bun install

Create a .env.local file:

ANTHROPIC_API_KEY=sk-ant-...         # Required
BRAVE_SEARCH_API_KEY=BSA...          # Optional — enables web search context

Start the dev server:

bun dev

Open http://localhost:3000.

This project is built in two stages.

Everything runs in the browser. Tasks are stored in IndexedDB, the scheduler uses setTimeout in an always-on tab, and AI calls are proxied through Next.js API routes with keys in .env.local. No backend, no accounts, no external infrastructure.

• Task CRUD with recurring schedules
• Client-side scheduling engine
• AI execution via Vercel AI SDK + Claude
• Web search context via Brave Search
• Activity feed, run history, markdown rendering
• Token usage tracking

Migrate to a proper backend with user accounts, server-side job scheduling, and persistent storage.

• Auth — user accounts and API key management per user
• Database — Supabase (Postgres) replacing IndexedDB
• Scheduling — Upstash QStash replacing client-side timers
• Multi-provider — OpenAI, xAI alongside Anthropic
• Notifications — browser push, email, Telegram
• Prompt templates — curated and user-created
• Tier system — free/pro/business with token caps and context limits
• Retry & fallback — automatic retry with fallback models on failure

Tunable constants live in src/lib/constants.ts:

export const MAX_OUTPUT_TOKENS = 4096;
export const DEFAULT_MODEL = "claude-sonnet-4-20250514";

MIT