Cost-effective sandboxes
for your agents.
Isolated Linux boxes for agents. Boot from your Dockerfile, pause for free, run in the EU.
Sign upPlans from $20 a month, after a 7-day free trial with $5 of credit.
What each sandbox gets.
Three things every agent needs. Few clouds promise them.
Bare metal we own
Our own servers, shared fairly. Most agent workloads never notice a neighbour.
See the plansPause costs nothing
Stop a sandbox and it keeps its disk: 32 GB on the $20 and $100 plans, 64 GB on the $500 plan. Resume it later and carry on. You pay only while it is running.
Keep as many paused sandboxes as you like, on every plan.
What survives a pause?EU compute
Sandboxes run in Germany and Finland. The fleet database, event store and template backups stay in the EU too.
Account, billing and analytics services operate from the US. They never see sandbox contents.
Read the privacy policyThree calls.
Start
Your Dockerfile becomes a template. One call turns it into a running machine with root.
import { Sandbox } from '@impello/sdk' const sandbox = await Sandbox.create('my-template')
Run
Give your agent a shell. Anything Linux can run, it can run: tests, builds, browsers, your own tools.
const result = await sandbox.commands.run('bun test') console.log(result.stdout)
Pause
Pause while the agent waits. Billing stops, the disk stays. Connect again later or kill it.
await sandbox.pause() // billing stops // later, from anywhere: const again = await Sandbox.connect(sandbox.sandboxId)
A browser your agent drives.
One of the things you can do with Impello. Start the browser template and your agent gets its own Chromium to work in.
- A real Chromium, running headful on its own display, not a headless build.
- Each profile keeps one fingerprint of its own, and the timezone follows the IP the browser browses from.
- Drive it over CDP with Playwright or Puppeteer, from your own machine.
- Watch it live, or take over, in noVNC on port 6080.
- 2 vCPU / 4 GB at $0.018 an hour. Pause it and the profile, cookies and logins stay on disk.
const sandbox = await Sandbox.create('browser', { network: { allowPublicTraffic: false }, }) const browser = await chromium.connectOverCDP( `https://${sandbox.getHost(9222)}`, { headers: { 'impello-traffic-access-token': sandbox.trafficAccessToken, } }, ) const page = browser.contexts()[0].pages()[0] await page.goto('https://example.com') // watch: https://<6080 host>/vnc.html
Simple pricing.
- 20 sandboxes at once
- 8 vCPU / 16 GB per sandbox
- 12 hours max length, 32 GB disk
- 5 parallel template builds
- Unlimited paused sandboxes
- 100 sandboxes at once
- 8 vCPU / 16 GB per sandbox
- 24 hours max length, 32 GB disk
- 20 parallel template builds
- Unlimited paused sandboxes
- 300 sandboxes at once
- 8 vCPU / 16 GB per sandbox
- 24 hours max length, 64 GB disk
- 40 parallel template builds
- Unlimited paused sandboxes
25 sandboxes · each runs 1% of the month
2 vCPU / 4 GB · $0.018 per hour · billed per second
Paused sandboxes are free. Plan credit expires monthly. Credit packs never expire.
| Free trial | $20 | $100 | $500 | |
|---|---|---|---|---|
| Price | $0 for 7 days | $20 / mo | $100 / mo | $500 / mo |
| Credit | $5, once | $20 / mo | $100 / mo | $500 / mo |
| Concurrent sandboxes | 2 | 20 | 100 | 300 |
| Max sandbox length | 1 h | 12 h | 24 h | 24 h |
| vCPU / RAM per sandbox | 4 / 8 GB | 8 / 16 GB | 8 / 16 GB | 8 / 16 GB |
| Disk per sandbox | 5 GB | 32 GB | 32 GB | 64 GB |
| Parallel template builds | 1 | 5 | 20 | 40 |
| Sandbox start limit, per minute | 5 | 12 | 30 | 65 |
| Sandbox start limit, per hour | 25 | 60 | 210 | 420 |
| Sandbox start limit, per day | 75 | 200 | 840 | 1,680 |
| Paused sandboxes | Unlimited | Unlimited | Unlimited | Unlimited |
| Region | EU | EU | EU | EU |
We're cheaper and
faster than most!
2 vCPU and 4 GB compared on a real job: the better-auth CI pipeline.
| Impello | $0.018 | 343 s | $0.0017 | free | — |
| Blaxel | $0.166 | 308 s | $0.0141 | free | 8.3× |
| E2B | $0.166 | 441 s | $0.0203 | free | 11.9× |
| Modal (V1 sandbox) | $0.238 | 341 s | $0.0225 | free | 13.2× |
| Modal (V2 sandbox) | $0.238 | 416 s | $0.0275 | free | 16.1× |
| Vercel Sandbox | $0.341 | 356 s | $0.0337 | $0.15 / GB | 19.7× |
| Fly.io Sprites | $0.315 | 520 s | $0.0455 | free | 26.6× |
| Railway Sandbox | $0.423 | 572 s | $0.0672 | $0.05 / GB | 39.3× |
Our own runs, 19 September 2026. Every price came off the provider's own page the same day. Click a column to sort the table. The method, every task time and every price basis →
What teams ask first.
Short answers here. Long ones in the legal pages.
Does a paused sandbox cost anything?
No. A paused sandbox holds its disk and costs nothing. Billing starts again when it resumes.
Do I share a CPU with other customers?
Yes, and so does every provider in the table above. A sandbox is a microVM on a shared host right across this market. Nobody sells you a whole machine at these prices. Where one advertises a "dedicated" size, the vCPUs are reserved while your box runs; the machine underneath is still shared.
Ours share servers we own, under a fair scheduler. Most agent workloads never notice. Sandboxes come in two sizes, 4 vCPU / 8 GB and 2 vCPU / 4 GB. Dedicated sandboxes for heavy workloads are on the roadmap.
How do you compare with other providers?
For the same 2 vCPU / 4 GB sandbox, the other providers we priced charge 9 to 24 times our hourly rate, and the ones we measured cost 8.3 to 39 times as much for one finished CI run.
The table above is our own benchmark, not a vendor leaderboard. We commit to being the lowest price on the market.
How can you charge this little?
We are a small team and we own the machines your sandboxes run on. We took no outside money, so there are no investors to pay back. That lets us charge close to what a sandbox actually costs us to run.
In our own benchmark we finish the same job within 12% of the fastest provider, and the others charge 3 to 8 times more an hour. You get the best mix of price and speed on the market.
What trade-offs did you make to charge this little?
Two, and you should hear them from us:
- Start time. Our sandboxes can take 1 to 3 seconds to launch, not in the 200 milliseconds the fastest providers advertise.
- Burst size. You cannot start 1,000 sandboxes in a minute. Our largest plan allows 65 starts per minute, and 300 sandboxes running concurrently.
Other providers have faster start times and support far higher bursting, but expect to pay 3 to 8 times more with them.
We find that most agent workloads don't need really fast start times (as agents still work slowly) and huge bursts either.
Where do sandboxes run?
In the EU, in Germany and Finland. The fleet database, event store and template backups are in the EU as well.
Payments, email and analytics operate from the US and never handle sandbox contents.
What survives pause and resume?
By default, everything. Memory and disk are snapshotted, so processes carry on where they stopped.
You can choose a disk-only snapshot instead; resuming that one reboots the machine from its disk.