Launch HN: Agent37 (YC F26) – Low-cost agent sandboxes for running agents

Agent37 showed up on Hacker News with a simple pitch: it wants to be the cheap place to run AI agents in sandboxes. Vishnu, the founder, says the company’s customers spend an average of $0.37 a month per sandbox, and that its setup comes in at under 10% of the cost of most sandbox providers.

The hook is the free plan. Vishnu says the team couldn’t afford to offer one at the start, when the company was bootstrapped and burning through his savings. Now it can, and he says more than 90% of new signups are free users hosting things like Hermes, OpenClaw and Claude Code on Agent37’s cloud.

The company’s cost story is really the product story. Agent37 says agent harnesses tend to use CPU in bursts and need extra RAM for browser and computer-use tasks, then quiet down again when the work is done. That pattern, the company argues, makes it possible to squeeze more value out of each machine if the infrastructure is tuned for it.

So it skips the hyperscalers and buys from providers such as Hetzner and OVH. It also uses gVisor to share host memory across sandboxes and speed up boot times, though that comes with limits: the gVisor sandboxes do not support things like Docker-in-Docker or Android emulators. For workloads that need more, Agent37 also offers full VM sandboxes at a higher price.

Vishnu frames the whole thing as a niche play. By focusing on a narrow class of sandbox workloads and optimizing around them, Agent37 says it can offer a better cost-performance tradeoff than broad, general-purpose providers. The company is also looking for developers building messaging agents to try pushing the platform’s limits and economics.