Agent37 Launches Sub-$1 AI Agent Sandboxes

YC F26 startup Agent37 has launched its host platform for running AI agents, promising costs under 10 percent of traditional sandbox providers. Founded by former Amazon senior software engineer Vishnu, who previously created Amazon's internal AI tools Diya and Ziya, the company reports that its users spend an average of $0.37 per month per sandbox. Agent37 also provides a free tier where infrastructure costs average just $0.05 to $0.10 per user monthly, attracting over 90 percent of its initial user base.

To achieve these price points, Agent37 bypasses hyperscale cloud providers like Amazon Web Services, which carry 80 to 90 percent margins, opting instead for bare-metal hosts from vendors such as Hetzner and OVH. The platform utilizes gVisor to share host memory across instances and deliver fast boot times without reserving dedicated virtual machine memory for every sandbox. Default environments provide 2 vCPU, 4 GB of RAM, and 4 GB of persistent disk. While this setup prevents workloads like Docker-in-Docker or Android emulation, Agent37 offers full virtual machines for workloads that require them.

The platform supports popular agent harnesses including Hermes, OpenClaw, Claude Code, OpenCode, Pi, and Grok. Vishnu noted the platform is roughly 2.8 times cheaper than competitor Boat and about 30 times less expensive than Cloudflare's sandboxes. For default model access without custom API keys, the service routes through OpenAI Luna under an agent37/default identifier. Alongside the sandboxes, Agent37 has open-sourced its Gateway API, which functions as a unified gateway converting various harness backends into an OpenAI-compatible API schema. For developers, this infrastructure provides an inexpensive environment to deploy continuous messaging agents and experimental workflows without facing hefty cloud compute bills.