Agent37 is a sandbox hosting platform for AI agents, whose core purpose is to provide an independent, isolated, and persistent runtime environment for each customer or each agent session. The platform claims costs as little as 95% lower than most sandbox providers, and promises to refund the difference if you find a cheaper sandbox.
Main capabilities:
- Multi-tenant sandbox isolation: create a sandbox per customer with a single call, and each sandbox has an independent file system, memory, and tool connections; files, memory, and credentials are isolated from each other until actively deleted.
- Full computing environment: run a real Linux machine, support installing software, executing scripts, and hosting applications, and provide root privileges. It also provides a browser and a full desktop that can be driven by the agent, while users can watch it work and take over when needed.
- Persistent disk: sandbox files are retained across sessions, sleep, and restarts; it automatically sleeps when idle and wakes on the next request.
- Rich agent ecosystem: supports hosting any agent framework, and custom Docker images can also be used; provides hosted models or bring your own API key; supports 1,000+ app connections through Composio; integrates with Perflo to purchase external capabilities per use, without vendor accounts and keys.
- Custom extensions: skills, prompts, and tools can be preinstalled through custom templates; supports setting cron schedules for agents to enable unattended operation.
- Lifecycle management: create, stop, restart, resize, update, and delete sandboxes; provides an authenticated web address for each agent; automatic backup and recovery; usage tracking and spending limit settings.
- Security and compliance: each sandbox is enforced with isolation by the gVisor kernel and has an independent file system and hard disk quota; the host has no public hostname, and all requests enter through the Cloudflare edge via TLS and require an API key; outbound traffic is restricted by a per-sandbox firewall to the public internet only; disks, backups, and metadata are encrypted at rest, and all connections are encrypted in transit. The platform does not collect prompts, files, or terminal sessions, data is not used to train AI models, and it provides SOC 2 Type I and readiness reports.
How to use: Users can paste the instructions provided by the platform into Claude Code, Cursor, or any coding agent to quickly build agent applications. The official recommendation is to read the documentation first:https://www.agent37.com/docs/llms-full.txt。平台按分钟计费：沙箱运行时计算计费，保留时磁盘计费；停机或睡眠后计算停止，删除后不再产生费用。采用预付费余额模式，无订阅、无套餐费、无最低消费；新账户绑定信用卡后获 $5 starting credit, no charge until top-up, sandbox sleeps when balance is exhausted. Model tokens, search, and integrations are metered from the same balance unless you bring your own key. Startups can receive up to $10,000 in credit (first $1,000, then matched spend).
Use cases: Used by over 100 companies for agent products. Examples include: Vision provides a separate machine for each real-time coding session; Boon provides Hermes agents for nonprofits to do grant fundraising; ETEC+ moved from manually creating agents to selling a white-label product with registration and billing; Perflo provides a persistent agent for each customer without hiring a team to run operations. Suitable for agent product developers who need multi-tenant isolation, persistent environments, and scalable infrastructure.
Limitations: The platform does not offer a free tier, and you need to bind a credit card to get the starting credit; outbound network access is limited to the public internet; the specific model support list and some integration details are subject to the official documentation.
