dafthunk

Dafthunk is an MIT-licensed visual workflow automation tool. You compose workflows in a React Flow editor, deploy them to Cloudflare, and trigger execution via HTTP, email, cron, queues, or manually. State persists in D1 and R2. The premise: infrastructure should be someone else's problem.

How it was built

This side project started as a sandbox. I teach software engineering and wanted to understand how AI-assisted coding changes web development (not in theory, in practice).

I began with Cursor and an assumption: perhaps 95% of the code could be generated and reviewed rather than written. That assumption held, and most of the edits happened through tab completion. I have since moved to Claude Code and now work in a somewhat post-IDE workflow. Going back feels awkward, though I suspect that says more about habits than objective improvement.

AI hallucinates regularly ... and so do I 🤪 In the end, managing AI agents feels a lot like managing people: be clear about what you want, delegate, trust but verify, and let go.

Why Cloudflare

The developer experience has been remarkably smooth, at least for this use cases. Cloudflare workers deploy in seconds. D1, R2, Workflows, and AI integrate without ceremony. Cloudflare's MCP servers and llms.txt support make AI-assisted development easier than I anticipated. I have not tried every alternative, but nothing else I have used comes close in terms of productivity.

A tentative observation

I noticed many developers, myself included, are resistant to AI-assisted coding and changes. It turns out we are not immune to the automation anxiety we have cheerfully inflicted on everyone else. Who knew, I did not expect my primary 2026 concern to become organizational psychology.

What's next

I plan to document progress here. If you work on serverless automation, or if you have tried similar experiments with AI-assisted development, I would welcome your perspective.