Multishot

Front screen when Multishot is opened
Sandboxes working in real time
Expanded display of Snake game created by one of the agents
Daytona-powered sandboxes running a Python version of Snake

We got tired of waiting for AI to generate code, only to realize it took the wrong approach. We envisioned an architecture that could run multiple AI agents at once with different strategies and pick the best solution. That's where Multishot comes in.

Multishot orchestrates 4 parallel Claude Code agents in Daytona-powered isolated cloud sandboxes. Each agent tackles your prompt with a different approach (thorough, minimal, edge-case focused, baseline). You watch the agents work in real-time, compare solutions side-by-side, select the winner, and download any generated files. Built-in Sentry monitoring auto-kills agents that error out, saving precious time, API credits, and compute power.

Our architecture is as follows: Electron for the desktop app, Daytona SDK for cloud-based sandboxes, Claude Code CLI running in each sandbox, Sentry for error tracking and auto-kill logic, React + Tailwind for the UI, and IPC handlers for backend/frontend communication.

These are the problems we ran into while debugging: Claude Code installation kept failing in sandboxes (curl → npm fallback saved us), Sentry v8 API changed - deprecated startTransaction() broke everything, file paths were all over the place (/workspace vs /home/daytona), require() doesn't work in renderer process - had to move logic to main process, and getting real-time terminal streaming from sandboxes.

We might have ran into lots of issues, but through tireless effort, we managed to debug and we are proud that we built a complete parallel AI orchestration system in one day, implemented Sentry for auto-kill logic that prevents wasted compute on broken agents, have smooth file downloads from cloud sandboxes to local machine, display a real-time terminal output for all 4 agents simultaneously, have a program that actually works and doesn't explode :-)

And all our debugging was actually a good thing, because we learned about Electron IPC architecture (main vs renderer process), Daytona SDK for managing cloud development environments, Sentry's capabilities for error monitoring and performance tracking, debugging remote sandbox environments, and never trusting default paths in containerized environments!

In the future, we plan to add more agent personalities for further customization (creative, pessimistic, optimistic), add a visual tool to compare code side-by-side, implementing saving winning solutions to a personal library, integration with more AI models (GPT, Gemini, etc.), and a collaborative mode, where you can share your agents racing against each other with your team.

We are proud to present Multishot, and we hope that you enjoy watching our demo as much as we enjoyed building it!