Developer Shukant Pal built Pretzel at the Google I/O hackathon on May 24, 2026, creating a live web-based music sequencer where an AI agent controls what everyone hears. The demo runs in any browser at pretzel.shukant.com and requires no signup.

What Happened

Pal set out to build a live DJ agent that users could describe feelings to, and the music would update in real time. Pretzel delivers a working version of that concept: all visitors share the same session, chat with the same AI agent, and hear the same music playing. Type a direction and the agent updates the sequencer live for everyone connected simultaneously. Pal posted the project on Hacker News with full technical notes on the architecture.

The backend runs on Talon, a Rust-based agent harness Pal has been developing that self-hosts on a micro VM using around 100MB of RAM. By connecting everyone to one Talon session, Pretzel keeps the experience collaborative and stateless from the user side. The creator describes the project as roughly halfway to the full vision, built in a single hackathon session.

Why It Matters

Pretzel is the simplest possible version of an idea that audio tool developers have been moving toward: natural-language control of a music sequencer in real time. Most current AI music tools take a prompt and return an audio file. Pretzel inverts that model: the AI agent is always running, always modifying the sequencer, and always responding to what people say.

This lands in the same space as recent live music diffusion research, but from a different direction. Where that work focuses on diffusion model inference speed, Pretzel focuses on agent architecture and collaborative interaction. Both point toward the same destination: AI music generation that responds to humans in the moment, not after a processing delay.

The rough edges are intentional. Pal describes it as 50 percent of the vision, built in a day by someone who is not an audio engineer. That constraint is useful context because the demo works despite it, which suggests the architecture holds up.

Key Details

  • Live demo: pretzel.shukant.com (no signup)
  • Built by: Shukant Pal at Google I/O hackathon, May 2026
  • Agent framework: Talon, a Rust runtime for long-lived agents (open source, launched May 2026)
  • Session model: All users influence the same sequencer and hear the same audio output
  • Status: Hackathon prototype, not a production tool

What to Do Next

Open pretzel.shukant.com in a browser and join the live session. Type a direction to the AI agent, such as a mood, tempo change, or style description, and watch the sequencer respond in real time. The experience takes under two minutes and gives a concrete sense of what AI-controlled collaborative music creation feels like at the interaction level. If you build audio tools or agent pipelines, the Talon source code on GitHub is worth reading for its approach to persistent, low-resource agent sessions running on minimal infrastructure.