Agent37, a YC F26 Company, Prices Agent Sandboxes at $0.37 for 57 Hours and Builds on gVisor

Agent37, a Y Combinator Fall 2026 company, says it runs AI agent sandboxes at under 10% of what most providers charge, and its own price page puts a 57-hour month at $0.37. The claim comes from founder Vishnu Kool's Launch HN post, which had 163 points on October 19. No funding figure appears on the pricing page or in the thread as read here.

Both the price and the comparison table are Agent37's own, and no independent party has run them.

What $0.37 buys

Agent37's pricing page lists prepaid, per-minute billing with no subscription. The rates are $0.0011 per vCPU-hour, $0.00096 per GB-hour of memory and $0.00012 per GB-hour of disk. The smallest configuration is 2 vCPU and 4 GB with a 4 GB persistent disk.

At those rates, the page says that size costs about $4.76 if it runs every minute of a month and $0.36 a month idle. The $0.37 figure is for roughly 57 hours of use. Kool says customers average $0.37 a month per sandbox.

The company's table prices the same 57 hours elsewhere.

1. Agent370.37 USD
2. AWS EC24.36 USD
3. Hetzner6.49 USD
4. Daytona and E2B9.44 USD
5. Cloudflare11.34 USD
6. Modal13.56 USD
7. Vercel Sandbox19.43 USD

Source: Agent37 pricing page (vendor-supplied), accessed 2026-10-19

Hetzner appears as a rival at $6.49 even though Kool says Agent37 rents from smaller hosts such as Hetzner and OVH. The gap between $6.49 and $0.37 is therefore packing, not the underlying hardware price.

How it gets the cost down

Kool lists four techniques: bursty agent workloads that share CPU, cheaper hosts than the hyperscalers, gVisor to share host memory instead of reserving a full VM's worth, and choosing machines with more RAM per dollar. Infrastructure costs about $0.05 to $0.10 per free user per month, he wrote.

The trade is in gVisor. Its documentation says it runs a user-space kernel, the Sentry, that does not pass system calls through to the host kernel, and that this "comes at the price of reduced application compatibility and higher per-system call overhead." Kool acknowledged that gVisor sandboxes cannot run Docker-in-Docker or Android emulators. Full VM sandboxes cost more.

What the thread pushed back on

A commenter, lasiotus, said gVisor is less secure than VMs and that the price comparison is not like for like. Another, ekorondy, wrote that "The host isolation everyone is comparing on price is not the hard problem," pointing to egress allowlists and brokered credentials. Commenters mparis and MikhailTal warned of a price race to the bottom. Kool replied that healthy margins let the company compete on price.

Two product gaps came up. An agent holding outbound polling connections cannot be woken by an inbound message, and Kool said a relay would fix it but is not built. Cold starts run 200 to 300 ms warm and a few seconds cold.

A free plan with two descriptions

Kool says most new signups are on the free plan. The pricing page describes a $5 starter credit, granted once a card is on file, and sandboxes sleep when it runs out. One commenter put the free limit at about 10 awake hours a week. The page lists no hours cap, so the two accounts differ.

A $5 credit against $0.05 to $0.10 of monthly cost per free user is 50 to 100 months of the cost Kool quotes, if usage stays small.

What would settle it

The company started as an OpenClaw wrapper in February and now hosts harnesses including Claude Code and Codex. Its startup program offers up to $10,000 in credits. The test is margin at scale, since the price is set against hosts it also rents from. For the same agent-tooling crowd, see CLI-Anything's 54,400 stars and Benchmark partners on rival conflicts. Agent37's next disclosure that would matter is a security write-up on its sandbox isolation, or a funding announcement.