Skip to content
Backed byY Combinator

Host any agent harness.

  • Hermes
  • OpenClaw
  • Claude Code
  • Codex
  • OpenCode
  • Grok
  • Custom harness

Give each customer an agent in its own isolated sandbox. Persistent files, browser access, and connections to Gmail, Slack, and other apps.

$1 free credit

Just want an agent for yourself? Create your free personal agent

Read the docs first: https://www.agent37.com/docs/llms-full.txt Build me a real estate assistant using Hermes.

Pricing

94% lower cost than most sandbox providers

Find a cheaper sandbox? We’ll refund the difference.

4 GB persistent disk

8% · 62 hrs
1 hourAlways running
Monthly estimate$0.73

Compute $0.37Disk $0.36

Monthly usage comparison

vs Agent37
  • Agent37$0.73
  • HetznerCX23 · EU$6.498.8×
  • AWS EC2c6a.large · US East$4.746.5×
  • Daytona$10.2714×
  • E2B$10.2714×
  • Cloudflare Sandbox2 vCPU · 6 GB$12.3416.8×
  • Modal$14.7520.1×
  • Vercel Sandbox$21.1328.8×

Unit rates

vCPU

$0.0011

per vCPU · hour

Memory

$0.00096

per GB · hour

Disk

$0.00012

per GB · hour

Compute is billed while your sandbox runs. Disk is billed while retained, awake or asleep.

Request

curl https://api.agent37.com/v1/instances \
  -H "Authorization: Bearer sk_live_..." \
  -H "Content-Type: application/json" \
  -d '{"template":"agent37-hermes"}'

Response

{
  "id": "ab12cd34ef",
  "status": "running"
}

One agent is an MVP.
Running 1,000 is an infra company.

Here's what you'd have to build.

Servers

Rent, harden, and scale servers as customers arrive.

Architecture

Every customer gets their own sandbox

One call provisions a sandbox per customer. Message it across as many sessions as you like. Its files, memory, and connected tools stay put until you delete it.

Agent 37 CloudYour appPOST/v1/instancesPOST/v1/responsesSandbox ArunningFiles, memory, and tools persist.HermesSandbox BrunningFiles, memory, and tools persist.OpenClawSandbox CrunningFiles, memory, and tools persist.Custom

Runs whatever you build with.

Agents

HermesOpenClawClaude CodeCodexOpenCodeGrok

Plus your own Docker image.

Models

ClaudeOpenAIGeminiGrok

Managed models, or bring your own key.

Apps

GmailSlackNotionGitHubGoogle Calendar

1,000+ app connections via Composio.

Custom skills & plugins

Preinstall your own skills, prompts, and tools with a custom template.

Streaming events

Stream text and tool activity as your agent works.

Scheduled runs

Give an agent a cron schedule; it runs unattended and you read the results.

Security

Every workload is treated as untrusted

We do not collect your prompts, files, or terminal sessions, and your data is never used to train AI models. SOC 2 Type I and readiness reports on request.

CONTROLS MONITOREDTYPE II IN PROGRESSSOC 2TYPE IView Trust Center

Isolation enforced by the kernel

Every sandbox is enforced by gVisor, with its own filesystem and a hard disk quota. Tenants can never see each other.

gVisor, built at Google

One door in

Hosts have no public hostname. Every request enters through Cloudflare's edge over TLS and needs your API key, unless you've explicitly made a port public.

Cloudflare global edge

Outbound to the internet only

Egress is firewalled per sandbox: the public internet and nothing else. Not the host, not private networks.

Deny by default

Encrypted at rest

Your sandbox's disk, backups, and metadata are encrypted at rest, and every connection is encrypted in transit.

At rest and in transit

Customer stories

What builders are shipping.

Agent products, customer workflows, and businesses built on Agent 37.

Wes Merrill

Cofounder @ Boon

We had tried for nearly 6 months to do this on our own, that took only a few days of setup and effort working with Agent37.

Boon gives nonprofits their own Hermes agents for grant fundraising.

Read the case study

Chris Nguyen

Founder @ ETEC+

Ideally I just want to sell, and not do any backend.

ETEC+ moved from creating agents by hand to selling a branded product with signup and billing.

Read the case study

FAQ

How does billing work?

You top up a prepaid balance and each sandbox spends it down per minute while it runs. No subscription, no plan fee, no minimum, and new accounts get a $1 starter credit once a card is on file, and nothing is charged until you top up. If the balance runs out, your sandboxes go to sleep until you top up again.

Is this a monthly subscription, or do I pay compute on top?

Neither. There is no subscription and nothing is added on top of it. You are charged per minute of uptime at your sandbox's own rate: a 2 vCPU 4 GB sandbox left running every minute of a month is $4.76, and the meter drops to its disk alone the minute it stops or sleeps. Delete it and you pay nothing. Model tokens, search, and integrations meter separately from the same balance, unless you bring your own keys.

What do I pay when an agent is idle?

Its disk alone, $0.36 a month for 4 GB. Compute stops the minute a sandbox stops or goes to sleep, and once you delete it you pay nothing.

Can I use my own model keys?

Yes, and that spend never touches your balance. Use our gateway instead and models, search, and integrations meter from the same balance, with a spend cap you set per sandbox.

How long can one agent run?

As long as you keep it. Nothing times out mid-task, and what your agent writes to disk is still there on the next call, weeks later.

Build the agent product your customers need.

Your harness, your workflows, your brand. Hosting, connections, models, and tools ready to build on.

Start Free

$1 free credit