Agent37 opens a free plan for persistent AI-agent sandboxes
The founder says lower-cost hosts and gVisor isolation make the free plan viable; the public site offers starting credit and metered usage.
A hosted home for each agent
Agent37’s founder announced a free plan on October 12 for a service that runs AI agents in separate, persistent sandboxes. The English product site offers a starting credit and metered compute and disk pricing. It presents a workflow in which a developer provisions a sandbox, sends an agent work through an API, and keeps its files and connected tools between sessions. The site lists support for harnesses including Hermes, OpenClaw, Claude Code and Codex, plus browser and terminal use.
Why the founder says it costs less
In the Launch HN post, the founder attributes the economics to matching agent workloads’ bursty CPU use, renting lower-cost hosts, and sharing host resources through gVisor. The claimed average costs and security properties are Agent37’s figures, not independent measurements. The founder also identifies a concrete tradeoff: gVisor sandboxes cannot run Docker-in-Docker or Android emulators. Agent37 says fuller virtual-machine sandboxes are available for such workloads at a higher price. Teams evaluating the service should compare their own runtime needs, isolation requirements and actual usage bills against those claims.