Cloudflare Sandbox vs Tensorlake

Both run each sandbox in its own microVM with its own kernel. Cloudflare Sandbox is a container inside the Workers platform. It sleeps when idle, wakes on the next request, and keeps files through snapshots or R2 that your code sets up. Tensorlake is a sandbox API whose named sandboxes suspend with their memory and running processes, can be copied into warm clones, and have /dev/kvm on CPU sandboxes.

The companies and products

Cloudflare Sandbox

Cloudflare Sandbox SDK runs untrusted or AI-generated code in Linux containers on Cloudflare's network, driven from a Cloudflare Worker.

Each sandbox is a Cloudflare Container backed by a Durable Object. Each container runs in its own Firecracker microVM with its own kernel and network. You build a linux/amd64 image, pick an instance type of up to 4 vCPU and 12 GiB, and reach the sandbox through your own Worker. The disk resets when the container stops, and your code keeps files with snapshots, automatic saves, R2 backups or R2 bucket mounts. It suits teams already on Workers who want sandboxes that sleep when idle. For code that only calls known methods, Cloudflare also offers Dynamic Workers, which run JavaScript in V8 isolates.

Tensorlake

Tensorlake is a San Francisco company that sells infrastructure for AI agents, with sandboxes as its main product.

Tensorlake Sandboxes are Firecracker microVMs that you create through an API and that boot in a few hundred milliseconds. CPU sandboxes have /dev/kvm, can resize CPU and memory while they run, and can suspend with their memory and running processes. You can copy a running sandbox into several warm clones. Around them Tensorlake sells pre-warmed pools, versioned Cloud Volumes, hosted Git and a bring-your-own-cloud option, aimed at teams running coding agents, evals and RL rollouts.

Headquarters
San Francisco

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. “43rd cheapest of 361” means 42 of the other 360 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 workloadCloudflare Sandbox: price rankTensorlake: price rankboat: price rank
Software factory
100 at once · 4 vCPU / 8 GB · 2,200 sessions of 8 h a month
Tensorlake fits a software factory: named sandboxes keep their files and memory through suspend, and CPU sandboxes have /dev/kvm. On Cloudflare your code saves snapshots, automatic saves or R2, and the disk is 20 GB at most. boat adds coding agents built in and saves the disk with no code.
192nd of 365, with limits
$1,543 a month
8th cheapest of 365
$1,315 a month
5th cheapest of 365
$634 a month
Always-on AI employees
50 machines 24/7 · 2 vCPU / 4 GB
Always-on assistants need their files to stay. Tensorlake keeps files and memory through suspend. On Cloudflare your code keeps files with snapshots or R2. boat saves the disk every minute and has no session cap once auto-stop is off (default 1 h).
43rd cheapest of 361
$2,188 a month
14th cheapest of 361
$1,730 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
Short code-interpreter calls suit both. Cloudflare bills CPU only while busy and wakes on the next request. Tensorlake boots in a few hundred milliseconds, bills CPU on active use, and publishes no start rate.
17th cheapest of 336
$144 a month
29th cheapest of 336
$250 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
Both fit sandboxes for your users. Cloudflare puts egress rules in Worker code, and Tensorlake takes allow and deny lists per sandbox. boat adds a per-sandbox usage API and per-member spend caps, with starts limited per plan.
19th cheapest of 349
$731 a month
13th cheapest of 349
$561 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
Tensorlake is built for RL rollouts: it runs up to 1,000 sandboxes at once on Pro and clones one warm state into many. Cloudflare starts new sandboxes from a saved snapshot. boat's start limits make large bursts a compromise.
147th of 349, with limits
$9,428 a month
2nd cheapest of 349
$5,637 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
60th cheapest of 359
$818 a month
35th cheapest of 359
$506 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
Docker builds need room. Cloudflare has 20 GB of disk at most and turns off iptables for Docker. Tensorlake runs Docker on its ubuntu-systemd image, with 10 to 100 GiB of disk. boat runs Docker with BuildKit and systemd, with 51 to 251 GB of disk.
303rd of 374, with limits
$704 a month
53rd cheapest of 374
$424 a month
2nd cheapest of 374
$86 a month
24/7 dev boxes
1 machine 24/7 · 4 vCPU / 8 GB
A dev box should keep its files. Tensorlake suspends a named sandbox after 10 idle minutes by default and keeps its state. Cloudflare starts a fresh disk after sleep unless your code restores a snapshot (beta, 30 days) or mounts an R2 bucket. boat keeps a stopped disk at least 14 days with no compute charge.
123rd cheapest of 371
$93 a month
202nd of 371, with limits
$250 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
Tensorlake ships an XFCE desktop with Firefox over VNC, with screenshot, mouse and keyboard calls. On Cloudflare you build a desktop into your image, or use the separate Browser Run service. boat has a 60 fps desktop with Chrome and computer tools.
207th of 367, with limits
$460 a month
10th cheapest of 367
$387 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
68th of 356, with limits
$233 a month
12th cheapest of 356
$250 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
102nd cheapest of 343
$286 a month
25th cheapest of 343
$250 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
168th of 327, with limits
$5,741 a month
123rd of 327, with limits
$3,484 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
25th cheapest of 351
$325 a month
8th cheapest of 351
$250 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

FeatureCloudflare SandboxTensorlakeboat
IsolationFirecracker microVM per container, own kernel and networkFirecracker microVMs with their own kernel; GPU sandboxes use gVisorFull VM with its own kernel, on a physical host dedicated to boat (AMD Ryzen 7 9700X or Ryzen 9 9950X); vCPUs are shared
Disk and resizeUp to 4 vCPU / 12 GiB / 20 GB disk per instance typeDisk 10 to 100 GiB, can only grow; CPU and memory resize on a running CPU sandboxDisk fixed per size, 51 to 251 GB; up to 16 vCPU / 32 GB; a resize needs a stop and resume
Files between sessionsDisk is fresh after sleep; keep files with snapshots (public beta, 30 days), scheduled automatic saves, R2 directory backups or a mounted R2 bucketNamed sandboxes suspend and keep files and memory; ephemeral ones end on idle timeout; Cloud Volumes keep files across sandboxesDisk saved every minute and at each stop, with no code; kept at least 14 days after a stop, longer as a saved snapshot
Memory snapshotsNo: snapshots save the filesystem, not memory or processesYes: suspend keeps memory and running processes and restores warmNo: the disk is kept, and a resume is like a reboot
Clone a running sandboxSnapshot or back up a running sandbox, then start new sandboxes from it; no one-call forkCopy a running or suspended sandbox into N warm clones with memoryFork a running sandbox in one call from its last 1-minute disk snapshot
Docker insidedocker:dind with iptables and IP forwarding turned offYes, on the tensorlake/ubuntu-systemd imageDocker with BuildKit and Compose, root and systemd
Nested virtualizationNot documentedYes: /dev/kvm in CPU sandboxes (not GPU sandboxes)Yes: /dev/kvm on every sandbox (Android emulators, VMs)
Desktop and computer useBuilt-in desktop removed in SDK 0.10.2; build one into your own image, or use the separate Browser Run service for headless browsersXFCE desktop with Firefox over VNC, with screenshot, mouse and keyboard calls; Chrome over CDP1920x1080, 60 fps desktop and Chrome, computer tools, embeddable stream; no GPU
Coding agentsGuides for Claude Code, Codex, OpenCode and Pi in your imageYou install your agent in the sandbox or drive it from outside; managed images are plain OS basesSeven installed (Claude Code, Codex, OpenCode and more), with a prompt, steer and events API
Keeping keys out of the sandboxYour outbound handler adds the token for any HTTP or HTTPS host; you write the codeStored secrets are injected into the sandbox as environment variablesOptional paid proxy ($1 a day while on, plus $1 per 1,000 requests); with it on, real keys of supported providers stay outside
Inbound trafficHTTP and WebSocket previews routed through your Worker; SSH; no raw TCP inbound or public IPHTTP, WebSocket and gRPC on exposed ports; SSH; raw TCP through authenticated client tunnels; no dedicated IPUp to 50 HTTPS URLs, token-gated by default; SSH; raw ports on a public IPv6 address
Outbound rulesallowedHosts deny-by-default list plus outbound handlers you write in Worker codeallow_out and deny_out lists by domain, IP or CIDR, or no internet at allNo provider-side allowlist; you set your own firewall as root
Idle handlingStops after inactivity and wakes on the next requestIdle timeout, 600 s by default, then suspend or terminateAuto-stop 1 h after create or resume by default (up to 30 days, or off on paid plans); no idle timer
GPUs and where it runsNo GPUs; several regions across Cloudflare's network, pin to a region or jurisdictionGPU sandboxes by request; EU data residency supported (no region list published), or your own cloud through BYOCNo GPUs; Germany, Finland and France; you cannot pick a region
Every feature Battleships tracks (43)
FeatureCloudflare SandboxTensorlakeboat
Full VM (own kernel)YesYesYes
Docker insideYesYesYes
Nested virtualizationNo documented wayYesYes
RootYesYesYes
SystemdNot publishedYesYes
Desktop GUINo documented wayYesYes
BrowserWith your own setupYesYes
Computer-use APINo documented wayYesYes
SSHYesYesYes
Persistent diskWith your own setupYesYes
Pause/resumeNo documented wayYesYes
Automatic snapshotsWith your own setupYesYes
Fork/cloneWith your own setupYesYes
Live fork (no pause)With your own setupYesYes
Memory forkNo documented wayYesNo 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 wayNo documented wayYes
Public IPv4No documented wayNo documented wayNo documented way
Egress allowlistYesYesNo documented way
Secret proxyYesNo documented wayYes
Secret proxy for any APIYesNo documented wayYes
Preinstalled agentsYesNo documented wayYes
Agent harness APIWith your own setupNo documented wayYes
MCP serverNo documented wayNo documented wayNo documented way
GPUNo documented wayYesNo 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 wayYesNo documented way
Open sourceYesNo documented wayNo documented way
Scoped API keysNot publishedNot publishedYes
Spend limitsNo documented wayNot publishedYes
Usage per sandbox (API)Not publishedNo documented wayYes
Audit logsYesNot publishedNo documented way
SSOYesYesNo documented way
SOC 2YesYesNo documented way
HIPAAYesYesNo documented way
EU data residencyYesYesNo documented way
Zero data retentionNot publishedYesYes

Performance

Cloudflare SandboxNot measured
TensorlakeNot measured
boat28.06 runs/s

Neither Cloudflare Sandbox nor Tensorlake is in StarSling's run of 3 October 2026 (run 37073680030), so we cannot say which is faster. In that run boat did 28.06 runs/s on the V8 Web Tooling Benchmark at 4 vCPU / 8 GB, second of 15. 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. In ComputeSDK's run of 6 November 2026, with 100 sandboxes started at once at each provider's default size, the median time to first command was 671 ms for Cloudflare and 1.55 s for Tensorlake. boat is not in ComputeSDK.

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. Cloudflare Sandbox and Tensorlake 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.

ScenarioCloudflare Sandbox, a monthTensorlake, 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
$39 · Workers Paid$250 · Pro$20 · $20/mo
One agent that never stops
4 vCPU / 8 GB, 50 GB disk, running 24/7 all month
$148 · Workers Paid$250 · Pro$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
$1,660 · Workers Paid$1,030 · Pro$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
$476 · Workers Paid$293 · Pro$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
$1,889 · Workers Paid · with limits$1,127 · Pro$8,800 · $2000/mo · with limits

Cloudflare Sandbox: You need Workers Paid ($5 a month minimum). CPU is billed only while busy, at $0.072 per vCPU-hour. Memory and disk are billed on the instance size in 10 ms steps until the container sleeps. 4 vCPU always comes with 12 GiB. Snapshot pricing is not published.

Tensorlake: Pro is $250 per billing cycle, which includes 25,000 CU ($250) of usage, then $0.01 per CU. CPU is billed on active execution, with a minimum of 20% of allocated vCPU. RAM and disk are billed on allocation while the sandbox runs, and a suspended sandbox pays $0.07 per GB-month of snapshot storage.

Tensorlake: Small users can buy prepaid Usage Credits packs of $5, $10 or $20 at higher rates instead of Pro; sandboxes stop when the balance hits zero. A free plan runs 1 sandbox of 1 vCPU / 1 GB for up to 2 hours. The tables price the Pro path.

Tensorlake: Plans cost $20, $100, $500 or $2,000 a month, and the whole plan price comes back as VM time for that month. A default 4 vCPU / 8 GB sandbox costs $0.036 an hour, billed per second while it runs, busy or idle. 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 at least 14 days. Saved snapshots above the included count ($0.056 a day each) and the AI proxy, if on, still bill while stopped.

Limits in those scenarios

Cloudflare keeps files between sessions only through snapshot, automatic-save or R2 code, and its local disk tops out at 20 GB. A Cloudflare account runs 1,500 vCPU at once by default, short of the 2,000 vCPU the 1,000-job burst needs, raised on request. Tensorlake fits all five scenarios on Pro, which allows 1,000 sandboxes at once; its disk is 10 to 100 GiB and an ephemeral sandbox ends after 10 idle minutes by default. boat fits four; the burst does not fit, because the $2,000 plan includes 90 starts a minute and 600 an hour, charges $0.03 to $0.07 for extra starts, and refuses starts above 270 a minute.

ScenarioCloudflare SandboxTensorlakeboat
One coding agent, workdaysNone for this scenarioNone for this scenarioNone for this scenario
One agent that never stopsNone for this scenarioNone for this scenarioNone for this scenario
50 agents in parallelNone for this scenarioNone for this scenarioNone for this scenario
Sandboxes for your usersNone for this scenarioNone for this scenarioNone for this scenario
1,000 short jobs at once
  • Workers Paid: max 1,500 vCPU running at once (needs 2,000)
None for this scenario
  • 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 Cloudflare Sandbox when

your app already runs on Workers and you want short sandboxes that sleep when idle, wake on the next request, and follow egress rules you write in code.

  • Programmable egress. Cloudflare can block every host except the ones you list, and run your own Worker code on each outbound request.
  • Sleeps when idle, wakes on request. Useful for many short, spiky sessions: a sandbox sleeps when quiet, wakes on the next request, and bills CPU only while busy.
  • Next to the rest of Workers. If your app already runs on Workers, sandboxes sit next to Durable Objects and R2, in a region you can pick.

Use Tensorlake when

agents must resume mid-process or fan out from one warm state, or when you need GPUs, live resize, outbound rules per sandbox or sandboxes in your own cloud account.

  • Memory snapshots and warm clones. Useful when an agent must pick up mid-process: Tensorlake saves a sandbox's memory and copies it into warm clones.
  • Large bursts with no published start rate. Pro runs up to 1,000 sandboxes at once, and named sandboxes keep their files and memory through suspend.
  • GPUs and your own cloud. Tensorlake offers GPU sandboxes by request and can run in your own AWS, GCP, Azure, CoreWeave, Nebius or bare-metal account.

Where boat fits

If your agents work for hours on builds, tests, emulators and a desktop, look at boat. It is a full VM that saves its disk with no code, runs coding agents you drive from outside, and with its optional paid proxy keeps supported model keys out of the machine. boat runs only in the EU, has no GPU and limits starts per plan.

Agents ready, keys kept outside

On both, you install the agent harness yourself. Tensorlake injects stored secrets into the sandbox the agent controls, and on Cloudflare you write and maintain the Worker code that adds every token. boat ships seven agents with a prompt, steer and events API, and an optional paid credentials proxy.

Reach it like a server

Cloudflare routes HTTP and WebSocket through your Worker, and Tensorlake reaches raw TCP only through authenticated client tunnels. boat gives up to 50 HTTPS URLs and raw ports on a public IPv6 address, so outside clients reach a database or game server directly.

A steady price for long sessions

One coding agent on workdays costs $20 a month on boat, $39 on Cloudflare and $250 on Tensorlake Pro. Cloudflare and Tensorlake bill CPU only while busy, so a mostly idle sandbox narrows the gap.

Use boat when your agents do real work for hours: builds, tests, Docker, emulators and a desktop, with coding agents built in, public ports and a VM that saves its disk with no code. Compare it one to one: boat vs Cloudflare Sandbox and boat vs Tensorlake.

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 Cloudflare Sandbox cheaper than Tensorlake?

It depends on the scenario. Cloudflare is lower for one workday agent ($39 against $250) and one always-on agent ($148 against $250), because Tensorlake Pro has a $250 minimum. Tensorlake is lower for 50 parallel agents ($1,030 against $1,660), 300 user sandboxes ($293 against $476) and 1,000 short jobs ($1,127 against $1,889). boat costs $20, $26, $317 and $200 for the first four, and $8,800 for the burst, because it bills extra starts.

Which is faster?

We cannot compare CPU: neither is in StarSling's run of 3 October 2026. For start time, ComputeSDK's run of 6 November 2026 measured a median of 671 ms for Cloudflare and 1.55 s for Tensorlake with 100 at once. boat did 28.06 V8 Web Tooling runs/s in StarSling's run, on an AMD Ryzen 9 9950X host; boat is not in ComputeSDK.

Which one keeps my files?

Both can. Tensorlake's named sandboxes suspend with their files, memory and running processes. Cloudflare keeps files with a snapshot (public beta, 30 days), scheduled automatic saves, an R2 backup or a mounted R2 bucket. boat saves the disk every minute and at each stop with no code, and keeps it at least 14 days after a stop.

Can I run Docker or Android emulators inside either one?

Docker, yes on both: Tensorlake on its ubuntu-systemd image, Cloudflare as docker:dind with iptables and IP forwarding turned off. Tensorlake CPU sandboxes also have /dev/kvm for emulators and VMs; Cloudflare does not document it. boat has Docker with BuildKit and /dev/kvm on every sandbox.

Which has GPUs?

Tensorlake offers GPU sandboxes by request, running on gVisor. Cloudflare Sandbox and boat have no GPUs.

Where do they run?

Cloudflare places containers in several regions of its network and lets you pin a region or jurisdiction. Tensorlake's docs say it supports EU data residency, and it can run in your own cloud through BYOC. boat runs on machines in Germany, Finland and France, and you cannot pick a region.

Try boat

Install the CLI, then run boat new.

7-day free trial

More comparisons