Boxes.dev: your laptop was never built for ten agents
Running five to ten Claude Code or Codex agents at once is normal now. Your laptop isn't. Close the lid and they stop. Ten of them trying to boot your app to test their changes will eat your RAM before lunch. Boxes.dev, which launched on YC on Nov 9, sells the obvious fix: a cloud computer for every agent.
The pitch is narrow on purpose. You paste a setup prompt into your local Claude Code or Codex, and the agent ports your dev environment to the cloud itself. After that, each agent gets its own machine with your full stack, runs the app, tests its own work end to end and sends screenshots. You can start agents from the desktop app, your phone, Slack or Linear, or port-forward to localhost and poke at the result yourself.
The line worth noticing is this one: "your dev environment shouldn't sell you tokens or dictate which harness you use." Boxes runs Claude Code and Codex unmodified, on your own subscriptions and API keys. That's a direct shot at the cloud-agent products that bundle a model, a harness and a markup together, and it's the same logic as Agent37 and the other YC sandbox startups we've covered: sell the machine, not the brain.
The founders, Nick and Drew, previously built Gem (YC S17) to over a $1B valuation. Their examples are big and self-reported: one customer ran a swarm of more than 200 agents on more than 100 machines to build complex software from scratch, with agents messaging each other, spawning new agents and machines, and organizing themselves. Others run "software factories" that trigger work through the API and output tested PRs. The launch link gives three free months of Pro.
Compare that with today's other headlines: a lab showing off an OS prototype built by over a thousand subagents, and a paper showing many small teams beat one giant team. Swarms are becoming an ordinary workload, and the bottleneck is moving from the model to the computers they run on. Whoever owns the box gets to watch every agent work, which makes this layer more strategic than it looks.
Source: https://www.ycombinator.com/launches/Wp5-boxes-dev-move-your-ai-coding-to-the-cloud and https://boxes.dev
← Back to all articles
The pitch is narrow on purpose. You paste a setup prompt into your local Claude Code or Codex, and the agent ports your dev environment to the cloud itself. After that, each agent gets its own machine with your full stack, runs the app, tests its own work end to end and sends screenshots. You can start agents from the desktop app, your phone, Slack or Linear, or port-forward to localhost and poke at the result yourself.
The line worth noticing is this one: "your dev environment shouldn't sell you tokens or dictate which harness you use." Boxes runs Claude Code and Codex unmodified, on your own subscriptions and API keys. That's a direct shot at the cloud-agent products that bundle a model, a harness and a markup together, and it's the same logic as Agent37 and the other YC sandbox startups we've covered: sell the machine, not the brain.
The founders, Nick and Drew, previously built Gem (YC S17) to over a $1B valuation. Their examples are big and self-reported: one customer ran a swarm of more than 200 agents on more than 100 machines to build complex software from scratch, with agents messaging each other, spawning new agents and machines, and organizing themselves. Others run "software factories" that trigger work through the API and output tested PRs. The launch link gives three free months of Pro.
Compare that with today's other headlines: a lab showing off an OS prototype built by over a thousand subagents, and a paper showing many small teams beat one giant team. Swarms are becoming an ordinary workload, and the bottleneck is moving from the model to the computers they run on. Whoever owns the box gets to watch every agent work, which makes this layer more strategic than it looks.
Source: https://www.ycombinator.com/launches/Wp5-boxes-dev-move-your-ai-coding-to-the-cloud and https://boxes.dev
Comments