Agent 37

Companies building AI agent products need isolated, per-customer compute environments to run agent harnesses, tools, files, and credentials, but setting up and operating this infrastructure themselves is costly and time-consuming. The page cites one founder who spent nearly six months trying to build this in-house before switching. Other customers describe wanting to focus on selling their product rather than running backend infrastructure or DevOps for a fleet of agent sandboxes. Standard cloud sandbox providers are also presented as expensive, with Agent 37's comparison table showing competitors like E2B, Daytona, Modal, and Vercel Sandbox costing 11 to 52 times more for equivalent usage.

Agent 37 runs cloud sandboxes that host agent harnesses such as Hermes, OpenClaw, Claude Code, Codex, OpenCode, Grok, or a customer's own Docker image, each isolated per customer with separate tools, files, and credentials. Billing is metered per minute of actual compute and disk usage with no subscription or plan fee, and sandboxes can auto-sleep when idle so only disk cost accrues. The platform includes networking (authenticated URLs, public ports, custom domains), lifecycle controls (start, stop, resize, restart), persistent storage with automatic backups across restarts, and observability for CPU, memory, disk, and logs. It also supports bringing your own model keys or using a managed gateway for models, 1,000+ app integrations via Composio, pay-per-call paid tools via Perflo, custom skill templates, and scheduled cron-style agent runs. Security controls include gVisor kernel-level isolation, Cloudflare-fronted access requiring API keys, per-sandbox firewalled egress, and encryption at rest and in transit, with SOC 2 Type I completed and Type II in progress.