CodeSandbox SDK vs Agent 37

CodeSandbox SDK gives you Firecracker VMs that hibernate with their memory, wake on the next HTTP request and fork from a warm state. Agent 37 hosts one instance per user with agents like Hermes, Claude Code and Codex built in, on gVisor sandboxes by default, billed per minute from a wallet.

The companies and products

CodeSandbox SDK

CodeSandbox SDK is an API from CodeSandbox, now a Together AI company, that starts sandboxed development VMs on the same infrastructure as the CodeSandbox editor.

Each sandbox is a Firecracker microVM built from a template or any Dockerfile. When it goes idle, it hibernates with a memory snapshot, and it resumes or forks from that snapshot in seconds. You drive it from a TypeScript SDK through terminals, commands, files and HTTPS hosts. Together's docs call it the deprecated offering and point new users to its successor, the together-sandbox SDK.

Agent 37

Agent 37 hosts AI agents, one isolated, always-on sandbox per user, behind one API.

You create an instance from a template (Hermes, OpenClaw, Claude Code, Codex, OpenCode, Pi, Grok or your own Docker image) and talk to it through a chat API, files, exec, SSH and HTTPS URLs. The default instance is a container enforced by gVisor. A VM type gives a full Linux machine with its own kernel for Docker, Kubernetes and Android emulators. It is built for companies that give each of their customers an agent.

Which use cases each one fits

For each use case, Battleships prices every sandbox provider on the same workload and sorts them from the cheapest monthly bill up. “32nd cheapest of 365” means 31 of the other 364 providers run that workload, with every feature it needs, for less. “With limits” means the provider runs it but misses something, so it is listed after every full match. Open a use case for its whole ranking and what each provider misses.

Use case and its workloadCodeSandbox SDK: price rankAgent 37: price rankboat: price rank
Software factory
100 at once · 4 vCPU / 8 GB · 2,200 sessions of 8 h a month
CodeSandbox SDK runs Docker on every VM and connects through its SDK, with no SSH documented. Agent 37 needs its VM type for Docker, and its entrypoint runs with no init system. boat gives every sandbox Docker, systemd, /dev/kvm and SSH.
32nd cheapest of 365
$3,731 a month
4th cheapest of 365
$551 a month
5th cheapest of 365
$634 a month
Always-on AI employees
50 machines 24/7 · 2 vCPU / 4 GB
Agent 37 ranks 1st here: one always-on assistant per user that can sleep when idle, keep its memory and wake on the next message. boat ranks 5th. CodeSandbox SDK also fits; it bills per started minute until a VM hibernates.
50th cheapest of 361
$4,811 a month
Cheapest of 361
$279 a month
6th cheapest of 361
$657 a month
Code interpreter bursts
100 at once · 1 vCPU / 2 GB · 200,000 sessions of 1 min a month
Both suit short code runs. CodeSandbox SDK wakes hibernated VMs with their memory, and Agent 37's default instances cost little per minute. boat's start limits make big bursts a compromise.
54th cheapest of 336
$642 a month
Cheapest of 336
$22 a month
189th of 336, with limits
$2,000 a month
Sandboxes for your users
300 at once · 2 vCPU / 4 GB · 30,000 sessions of 23 min a month
Agent 37 ranks 1st and boat 2nd. Agent 37 allows up to 1,000 awake instances after $1,000 of top-ups, and sleeping ones do not count. CodeSandbox SDK's Scale plan stops at 250 VMs at once. boat allows 100 to 2,000 at once by plan, within per-plan start limits.
140th of 349, with limits
$1,707 a month
Cheapest of 349
$79 a month
3rd cheapest of 349
$209 a month
RL rollouts
1,000 at once · 2 vCPU / 4 GB · 1,000,000 sessions of 3 min a month
CodeSandbox SDK's forks can share memory, so many rollouts can branch from one running state. Agent 37 and boat fork the disk, and the fork boots fresh.
341st of 349, with limits
$3,861 a month
9th of 349, with limits
$363 a month
39th of 349, with limits
$2,000 a month
Agent evals
250 at once · 4 vCPU / 8 GB · 15,000 sessions of 20 min a month
77th cheapest of 359
$1,335 a month
Cheapest of 359
$58 a month
10th cheapest of 359
$200 a month
CI runners
40 at once · 8 vCPU / 16 GB · 6,000 sessions of 12 min a month
Both run Docker builds, Agent 37 only on its VM type. Jobs that need /dev/kvm work on Agent 37's VM type and on every boat sandbox; CodeSandbox SDK does not document it.
96th cheapest of 374
$860 a month
227th of 374, with limits
$96 a month
2nd cheapest of 374
$86 a month
24/7 dev boxes
1 machine 24/7 · 4 vCPU / 8 GB
316th of 371, with limits
$240 a month
3rd cheapest of 371
$12 a month
19th cheapest of 371
$26 a month
Computer-use agents
50 at once · 4 vCPU / 8 GB · 20,000 sessions of 15 min a month
Agent 37 ranks 1st and boat 8th. Agent 37 adds a noVNC desktop through a template. On CodeSandbox SDK you add a VNC desktop to your own Dockerfile. boat streams a built-in 60 fps desktop with Chrome.
54th cheapest of 367
$1,236 a month
3rd cheapest of 367
$120 a month
7th cheapest of 367
$200 a month
Browser-use agents
50 at once · 2 vCPU / 4 GB · 30,000 sessions of 10 min a month
125th of 356, with limits
$741 a month
47th of 356, with limits
$29 a month
7th cheapest of 356
$200 a month
Scraping & bots (startup scale)
50 at once · 1 vCPU / 2 GB · 100,000 sessions of 3 min a month
119th cheapest of 343
$871 a month
Cheapest of 343
$32 a month
6th cheapest of 343
$90 a month
Mass scraping & bots (undetectable)
1,000 at once · 1 vCPU / 2 GB · 3,000,000 sessions of 2 min a month
194th of 327, with limits
$7,390 a month
39th of 327, with limits
$648 a month
73rd of 327, with limits
$1,800 a month
Parallel red teaming
200 at once · 2 vCPU / 4 GB · 5,000 sessions of 30 min a month
Agent 37 ranks 1st for many cheap parallel attempts. All three can fork: CodeSandbox SDK with memory, Agent 37 and boat with the disk only.
29th cheapest of 351
$505 a month
Cheapest of 351
$43 a month
3rd cheapest of 351
$200 a month
GPU eval jobs
10 at once · 8 vCPU / 32 GB · 500 sessions of 30 min a month
Not ranked: misses a requirementNot ranked: misses a requirementNot ranked: misses a requirement
macOS agents
5 at once · 4 vCPU / 8 GB · 1,000 sessions of 20 min a month
Not ranked: misses a requirementNot ranked: misses a requirementNot ranked: misses a requirement

Key features

FeatureCodeSandbox SDKAgent 37boat
IsolationFirecracker microVM with its own kernelgVisor sandbox by default; full VM as an opt-in instance typeEvery sandbox is a full VM with its own kernel, on hosts dedicated to boat (AMD Ryzen 7 9700X and Ryzen 9 9950X); vCPUs are shared
Docker insideDocker and root on every VM, built from any DockerfileOn the VM type, at 2.5x the default compute rate; the default gVisor type cannot run a Docker daemonDocker, systemd and root in every sandbox
Nested virtualizationNot documented/dev/kvm on the VM type, with a ready Android template/dev/kvm in every sandbox; Android emulators work
Shell accessTerminals, commands and files through the SDK; SSH is not documentedSSH through an authenticated WebSocket tunnel, plus exec and filesBuilt-in SSH, plus exec and files over the API
Memory across sleepHibernates with a memory snapshot after an idle timeout of up to one day; resumes in about 0.5 to 2 sOpt-in auto-sleep (default 15 minutes) checkpoints memory; processes resume when the restore worksDisk saved every minute and on stop; memory is not saved, so a resume is like a reboot
Wake on requestYes, an HTTP request to a host wakes itYes, wakes on the next requestNo; you resume a stopped sandbox by API, CLI or SDK
ForkFork a hibernated sandbox from its snapshot, or a running one sharing its memoryFork any running, stopped or sleeping instance (disk copy, the fork boots fresh); nightly backups in 7 slots plus one on-demandFork a running sandbox from its last 1-minute snapshot (disk only, the fork boots fresh); saved snapshots as templates
Biggest machine16 cores / 32 GB on Scale, up to 64 cores / 128 GB on Enterprise4 vCPU on every type16 vCPU / 32 GB on the $100 plan and above
Custom imagesTemplates from any DockerfileYour own linux/amd64 Docker image as a templateStart from boat's image or a saved snapshot; not your own Docker image
Built-in agentsYou set it up: install your agent in the template and run it through the commands APIHermes, OpenClaw, Claude Code, Codex, Grok, OpenCode and Pi, with a chat API per instanceSeven harnesses, including Claude Code, Codex, pi and OpenCode, with a prompt, steer and events API
Desktop and browserHeadless Chromium template; you can install a VNC desktop in your own DockerfileYou set it up: noVNC desktop through a template build, or the Ubuntu VM template1920x1080, 60 fps desktop with Chrome and computer-use tools in every sandbox
Keys kept out of the sandboxSecrets go in as environment variables; no credentials proxy documentedManaged LLM, search and Composio through a per-instance token; your own keys go in as env varsOptional AI credentials proxy ($1 a day plus $1 per 1,000 requests) for model providers and HTTPS APIs you register; other secrets still enter
Inbound accessHTTPS host per port, public by default, signed tokens for private sandboxes; raw TCP not documentedHTTPS URL per port, signed URLs, custom domains; no public IPUp to 50 HTTPS URLs per sandbox, token-gated by default; SSH; raw ports over public IPv6; the IP changes at each resume
RegionsA country hint steers scheduling, for example to Europe, but placement is not guaranteedDedicated servers in the United StatesGermany, Finland and France; you cannot pick a region; no US or Asia region
Every feature Battleships tracks (43)
FeatureCodeSandbox SDKAgent 37boat
Full VM (own kernel)YesNo documented wayYes
Docker insideYesNo documented wayYes
Nested virtualizationNo documented wayNo documented wayYes
RootYesYesYes
SystemdNot publishedNot publishedYes
Desktop GUIWith your own setupWith your own setupYes
BrowserWith your own setupYesYes
Computer-use APINo documented wayNo documented wayYes
SSHNo documented wayYesYes
Persistent diskYesYesYes
Pause/resumeYesYesYes
Automatic snapshotsYesYesYes
Fork/cloneYesYesYes
Live fork (no pause)YesYesYes
Memory forkYesNo documented wayNo documented way
Custom image (Docker or snapshot)YesYesYes
≥24 h sessionsYesYesYes
Idle auto-stopYesYesNo documented way
Wake on requestYesYesNo documented way
HTTPS preview URLsYesYesYes
Raw TCP inboundNo documented wayNot publishedYes
Public IPv4No documented wayNo documented wayNo documented way
Egress allowlistNo documented wayNo documented wayNo documented way
Secret proxyNo documented wayYesYes
Secret proxy for any APINo documented wayNo documented wayYes
Preinstalled agentsNo documented wayYesYes
Agent harness APINo documented wayYesYes
MCP serverNo documented wayNo documented wayNo documented way
GPUNo documented wayNo documented wayNo documented way
WindowsNo documented wayNo documented wayNo documented way
Arm64No documented wayNo documented wayNo documented way
Self-hostingNo documented wayNo documented wayNo documented way
BYOCNo documented wayNo documented wayNo documented way
Open sourceNo documented wayNo documented wayNo documented way
Scoped API keysNot publishedNot publishedYes
Spend limitsNot publishedYesYes
Usage per sandbox (API)No documented wayYesYes
Audit logsNot publishedNot publishedNo documented way
SSOYesNo documented wayNo documented way
SOC 2YesYesNo documented way
HIPAANo documented wayNo documented wayNo documented way
EU data residencyNo documented wayNo documented wayNo documented way
Zero data retentionNot publishedNot publishedYes

Performance

CodeSandbox SDKNot measured
Agent 37Not measured
boat28.06 runs/s

CodeSandbox SDK and Agent 37 are not in StarSling's run of 3 October 2026 (run 37073680030), so we cannot say which is faster at CPU work. In that run boat did 28.06 V8 Web Tooling Benchmark runs per second at 4 vCPU / 8 GB, second of 15. Every boat sandbox in that run ran on an AMD Ryzen 9 9950X host; boat also places sandboxes on Ryzen 7 9700X hosts and does not promise a CPU model. For start time, ComputeSDK measured CodeSandbox SDK at 11.76 s median with 100 at once (6 November 2026), at its default size. Agent 37 and boat are not in ComputeSDK. Agent 37's default instances share host resources, and its Performance type runs on dedicated cores.

V8 Web Tooling Benchmark at a 4 vCPU / 8 GB target, runs per second, higher is better. Source: StarSling's run of 3 October 2026 (run 37073680030), median of 5 sandboxes. Providers ran on different hardware in different regions. StarSling gives a shared rank when the run cannot separate two providers. Every boat sandbox in this run ran on an AMD Ryzen 9 9950X host. boat also places some sandboxes on AMD Ryzen 7 9700X hosts, which this run did not measure. CodeSandbox SDK and Agent 37 are not in the public benchmarks we use, so we cannot say which is faster.

Pricing in five scenarios

Each bill is priced from the providers' public price lists by the open Battleships engine, with the cheapest plan that runs the scenario. Start limits, per-start fees and plan minimums are in the price, boat's included.

ScenarioCodeSandbox SDK, a monthAgent 37, a monthboat, a month
One coding agent, workdays
4 vCPU / 8 GB, 50 GB disk, 8 hours a day, 22 days a month, files kept between days
$193 · Scale$6 · Wallet (any top-up)$20 · $20/mo
One agent that never stops
4 vCPU / 8 GB, 50 GB disk, running 24/7 all month
$341 · Scale$12 · Wallet (any top-up)$26 · $20/mo
50 agents in parallel
50 machines of 4 vCPU / 8 GB at once, 8-hour sessions, 1,100 sessions a month, Docker inside
$2,500 · Scale$329 · Wallet ($50 lifetime top-ups)$317 · $20/mo × 3 seats
Sandboxes for your users
2 vCPU / 4 GB, 30,000 sessions of 10 minutes a month, up to 300 at once
$815 · Scale · with limits$33 · Wallet ($1,000 lifetime top-ups)$200 · $100/mo × 2 seats
1,000 short jobs at once
2 vCPU / 4 GB, 1,000 at once, 3 minutes each, 200,000 jobs a month
$1,632 · Scale · with limits$73 · Wallet ($1,000 lifetime top-ups) · with limits$8,800 · $2000/mo · with limits

CodeSandbox SDK: CodeSandbox SDK sells VM credits at $0.01486 each, and a 4 vCPU / 8 GB VM burns 20 an hour, about $0.30. Billing is per started minute while the VM runs, including idle time before it hibernates. A hibernated VM costs nothing. The Scale plan adds a $170 monthly platform fee that is not usage credit.

Agent 37: A prepaid wallet with no plan fee, metered per minute: $0.80 per vCPU and $0.70 per GB of RAM per month while running, and $0.09 per GB of disk per month even while stopped or asleep. Performance compute costs 4x and VM compute 2.5x the default rate. New accounts get a one-time $5 credit after adding a card.

Agent 37: Plans cost $20, $100, $500 or $2,000 a month, and the whole plan price comes back as VM time for that month; it does not carry over. A default 4 vCPU / 8 GB sandbox costs $0.036 an hour, billed per second while it runs. The $20 plan includes 12 starts a minute, 60 an hour and 200 a day; extra starts cost $0.03 to $0.07. A stopped sandbox has no compute charge and keeps its disk for at least 14 days; saved snapshots above the included count and the AI credentials proxy still charge.

Limits in those scenarios

CodeSandbox SDK's Scale plan allows 250 VMs at once, so 300 user sandboxes and 1,000 jobs at once need an Enterprise plan. It gives each VM 20 GB of disk and publishes no price for more disk or for egress. Together's docs call this SDK deprecated; the last release was 2.4.2 in December 2025. Agent 37 caps awake instances by lifetime top-ups (3, then 10, 50, 200 and 1,000 at $1,000); sleeping and stopped instances do not count. Every Agent 37 type stops at 4 vCPU, and its 50-agent scenario needs Docker, so it is priced on the VM type at 2.5x. On boat the 1,000-job burst hits the start limits: the $2,000 plan charges for starts above 90 a minute, refuses them above 270, and never queues a start.

ScenarioCodeSandbox SDKAgent 37boat
One coding agent, workdays
  • disk beyond 20 GiB: price unknown
  • egress price not published
  • egress price not published
None for this scenario
One agent that never stops
  • disk beyond 20 GiB: price unknown
  • egress price not published
  • egress price not published
None for this scenario
50 agents in parallel
  • egress price not published
  • egress price not published
None for this scenario
Sandboxes for your users
  • Scale: max 250 concurrent
  • egress price not published
  • egress price not published
None for this scenario
1,000 short jobs at once
  • Scale: max 250 concurrent
  • egress price not published
  • shared vCPUs on this tier: heavy, sustained CPU use can be slowed by other tenants
  • egress price not published
  • 2,400 starts/day included, paid above that (needs ~6,667)
  • 600 starts/hour included, paid above that (needs ~1,000)
  • starts above 90/min are paid, and starts above 270/min fail with 429

Which one should you use?

Use CodeSandbox SDK when

your sandboxes sleep between visits and should wake on the next HTTP request with their memory intact, or when you fork many runs from one warm state, and you accept that Together calls this SDK deprecated.

  • Hibernation with memory and wake on request. Good for previews and code runs used in bursts: an HTTP request wakes a hibernated VM with its processes still running.
  • Forks that share memory. Good for branching many runs from one warm state, such as rollouts or test variants, without booting each copy again.
  • Docker everywhere, bigger machines. Every VM runs Docker from any Dockerfile. Scale goes to 16 cores and 32 GB, and Enterprise to 64 cores and 128 GB.

Use Agent 37 when

you host one mostly idle assistant per customer in the US, from a ready agent template, and want a small per-minute bill with memory kept while it sleeps.

  • Built for one agent per customer. Pick a template like Hermes, OpenClaw, Claude Code or Codex and talk to each user's instance through a chat API.
  • Small bills for mostly idle assistants. A prepaid wallet billed per minute with no plan fee. One agent that never stops costs $12 a month, and idle instances can sleep with their memory kept.
  • Your Docker image and dedicated cores. Build templates from your own Docker image, and pick dedicated cores on the Performance type, close to US users.

Where boat fits

Both are good at sandboxes that sleep with their memory and wake on request. If your agents instead work for hours on installs, builds, Docker, emulators and a browser, boat gives every sandbox a full VM with SSH, systemd, a desktop and coding agents set up. boat does not keep memory across a stop, and it runs only in Germany, Finland and France.

A full machine by default

CodeSandbox SDK documents no SSH and no /dev/kvm, so Android emulators are not a supported path. Agent 37's default gVisor type cannot run Docker; its VM type can, at 2.5x the compute rate, and its entrypoint runs with no init system. boat has Docker, systemd and /dev/kvm in every sandbox.

A desktop for computer use

On CodeSandbox SDK you add a VNC desktop to your own template. Agent 37 adds a noVNC desktop through a template build. boat streams a 60 fps desktop with Chrome and computer-use tools in every sandbox.

A bigger disk, saved every minute

CodeSandbox SDK gives each VM 20 GB, with no published price for more. Agent 37 bills $0.09 per GB of disk a month, also while asleep. A default boat sandbox has 51 GB, saved every minute and at each stop, and kept at least 14 days while stopped.

Use boat when your agents do real work for hours: installs, builds, Docker, emulators and a browser on a full VM with SSH and systemd, a disk saved every minute, and forks of running machines. Compare it one to one: boat vs CodeSandbox SDK and boat vs Agent 37.

boat's own limits
  • Plans start at $20 a month. The plan price comes back as VM time for that month. Unused time does not carry over. Source
  • Starts (create, fork, resume): the $20 plan includes 12 a minute, 60 an hour and 200 a day. The $2,000 plan includes 90 a minute, 600 an hour and 2,400 a day. Extra starts cost $0.03 to $0.07. A start fails with 429 above 3 times the included starts per minute, or 6 times per hour or per day. Source
  • A stopped sandbox has no compute charge. boat keeps its disk for at least 14 days. A saved snapshot keeps it longer; extra saved snapshots and the AI credentials proxy are paid. Source
  • Auto-stop defaults to 1 hour. You can turn it off on a paid plan. The trial allows 2 hours at most. Source
  • boat saves the disk every minute and at each stop. Memory, running programs and /tmp are not saved: a resume is like a reboot. A machine failure loses writes after the last good snapshot. Source
  • Machines run in Germany, Finland and France only. Snapshot copies can be stored outside the EU. Source
  • Public IPv6 and outbound IPv4. No dedicated IPv4. The IP changes at each resume. Source
  • Hosts are AMD Ryzen 7 9700X and Ryzen 9 9950X. vCPUs are shared. No GPU. You cannot boot your own Docker image. Source

Questions people ask

Is CodeSandbox SDK cheaper than Agent 37?

No. Agent 37 comes out lower in every scenario we priced. One agent that never stops is $12 a month on Agent 37 against $341 on CodeSandbox SDK, and 50 agents in parallel are $329 against $2,500. boat comes to $26 and $317 for the same two.

Which is faster, CodeSandbox SDK or Agent 37?

We cannot say: neither is in the public CPU benchmarks we use. Agent 37 sells a Performance type on dedicated cores for steady speed. boat did 28.06 V8 Web Tooling Benchmark runs per second at 4 vCPU / 8 GB in StarSling's run of 3 October 2026, on an AMD Ryzen 9 9950X host.

Can CodeSandbox SDK and Agent 37 run Docker?

CodeSandbox SDK runs Docker on every VM and builds from any Dockerfile. Agent 37 runs Docker only on its VM type, which bills compute at 2.5x; the default gVisor type cannot run a Docker daemon. boat runs Docker in every sandbox.

Do both keep memory when a sandbox sleeps?

Yes. CodeSandbox SDK hibernates with a memory snapshot, and Agent 37 checkpoints memory on auto-sleep, and both wake on the next request. boat keeps the disk, but a resume is like a reboot.

Is CodeSandbox SDK still maintained?

CodeSandbox is now part of Together AI. Together's docs call this SDK deprecated and name together-sandbox, which your organization must ask Together to enable, as its successor. The last SDK release was 2.4.2 in December 2025; self-serve sign-up still runs through CodeSandbox.

Where do they run?

Agent 37 runs on dedicated servers in the United States. CodeSandbox SDK takes a country hint, for example to a cluster in Europe, but does not guarantee placement. boat runs sandboxes on machines in Germany, Finland and France, and you cannot pick a region; snapshot copies can be stored outside the EU.

Try boat

Install the CLI, then run boat new.

7-day free trial

More comparisons