ComputeSDK: one API for sandboxes across 38 providers

TL;DR: ComputeSDK gives you one TypeScript API and one REST API for spinning up remote sandboxes across 38 providers like E2B and Modal, so you can swap vendors without rewriting your application. The same project publishes free weekly benchmarks on sandbox cold starts, storage, and browser throughput, with the raw JSON sitting in a public repo.

What it is

Two things live under the ComputeSDK name, and they were built to feed each other.

The first is an MIT-licensed TypeScript SDK. You configure a provider once, then call the same create, runCommand, and destroy methods no matter which vendor sits behind them. Provider packages ship separately, so @computesdk/e2b and @computesdk/modal each wrap one vendor’s API and normalize it to an identical interface. The core package has 282 GitHub stars and was pushed the day before this post.

The second is the benchmark platform around it. ComputeSDK runs the same workload against every provider on a schedule, scores the results, and publishes the rankings every Friday. The reasoning is straightforward: picking a sandbox vendor is something you need a measurement for, not something a marketing page answers.

The company behind it is Snelling, LLC, which also sells Stealth, the same suite run daily and privately for a single provider.

The benchmark side

Here is the current sandbox burst TTI table, the median time-to-interactive when launching many sandboxes concurrently, pulled straight from the public leaderboard:

Provider	TTI (median)	Composite	Success
isorun	59 ms	99.36	100%
miosa	73 ms	98.59	100%
promigence	193 ms	97.96	100%
archil	210 ms	97.62	100%
arker	309 ms	96.81	100%
vercel	417 ms	95.27	100%
e2b	1334 ms	85.79	100%
cloud-run	2161 ms	69.77	100%
sail	3857 ms	61.42	100%
sandbox0	8075 ms	6.82	59%
codesandbox	38356 ms	0	99%

The fastest provider is roughly 650 times quicker than the slowest one on this table. E2B lands in the middle at 1.3 seconds, behind the leader but well ahead of the bottom half.

Storage and browser benchmarks follow the same format:

Storage provider	Upload	Download	Composite	Success
azure-blob	146 ms	219 ms	96.19	100%
archil	431 ms	134 ms	97.13	100%
aws-s3	603 ms	164 ms	96.25	100%
cloudflare-r2	865 ms	535 ms	86.14	100%
neon	983 ms	152 ms	94.04	100%
gcs	0 ms	0 ms	0	0%

Browser provider	Actions per second	Composite	Success
tilion	7.39	85.98	99%
browserbase	4.25	72.11	99%
steel	3.49	67.93	99%
hyperbrowser	2.79	31.68	66%

Two rows deserve a closer look. The gcs row scoring zero is a failure, not a fast result, and hyperbrowser posting 66% success means roughly a third of its runs broke. Publishing failure rates next to the speed numbers is rarer than it should be, and it is the part that makes the rest of the table worth trusting.

Getting it running

Install the core SDK plus whichever provider you want to start with:

npm install computesdk @computesdk/e2b

Then configure the provider and run something inside it:

import { compute } from 'computesdk';
import { e2b } from '@computesdk/e2b';

compute.setConfig({
  provider: e2b({ apiKey: process.env.E2B_API_KEY }),
});

const sandbox = await compute.instances.create();

const result = await sandbox.runCommand('python -c "print(\'Hello World!\')"');
console.log(result.stdout); // "Hello World!"

await sandbox.destroy();

What matters here is what stays fixed. compute.instances.create() does not change when you swap e2b for modal or vercel. Filesystem operations, command execution, and interactive PTY terminals are normalized across the roster, and the SDK exposes detached processes for commands that outlive the request that started them.

Using it over HTTP

You do not need Node in the picture. The platform runs a REST API at https://platform.computesdk.com/api/v1 with bearer auth, and the OpenAPI spec is served at https://www.computesdk.com/openapi.json.

Create a sandbox, then run a command inside it:

curl -X POST https://platform.computesdk.com/api/v1/sandboxes \
  -H "Authorization: Bearer $COMPUTESDK_KEY" \
  -H "Content-Type: application/json" \
  -d '{"label": "build-42", "timeoutMs": 300000}'

curl -X POST https://platform.computesdk.com/api/v1/sandboxes/$SANDBOX_ID/commands \
  -H "Authorization: Bearer $COMPUTESDK_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command": "python -m pytest -q"}'

The create request is where the interesting options live. providerOrder pins or orders placement across the roster, image selects a base image, snapshotId or disk restores prior state, and maxPrice turns the request into a limit order on the Market rather than a normal placement.

There is also an MCP server at https://platform.computesdk.com/mcp, spoken over Streamable HTTP with OAuth 2.0, so agents can create sandboxes without going through your own tooling. The website itself answers Accept: text/markdown on every page, which makes it unusually easy to pull into an agent’s context.

What else is in the box

Actions runs git workflows on ComputeSDK sandboxes, so a repository can define CI steps that execute on remote sandboxes instead of a shared runner. The Market is the other side of the same arrangement: providers list idle capacity at their own listed price, and buyers pay the ask.

Pricing splits cleanly along the same line. Public benchmarks and platform access are free. Stealth is the paid product, where one provider gets the full suite run daily and privately against pinned regions.

What to watch for

The composite score is a ComputeSDK formula, not an industry standard. The raw JSON is public in the computesdk/benchmarks repo, so you can check the weights and recompute the rankings yourself, but any ordering you see published is still one company’s weighting.

Provider churn is the bigger operational risk. The roster shifts, and a provider sitting at the top of the table today can be gone or repriced within a quarter. That cuts both ways: it is an argument for using the SDK’s abstraction instead of a vendor’s API directly, and an argument against hard-wiring yourself to a leaderboard position.

On the SDK side, provider packages are versioned separately from the core package, so mismatched versions are possible. Nothing in the repository prevents it, which means pinning them together is worth doing when it matters.

Source and Accuracy Notes

• Project page: https://www.computesdk.com/ (verified via direct visit)
• Source repository: computesdk/computesdk (MIT, TypeScript, 282 stars) and computesdk/benchmarks (MIT)
• License: MIT, verified from the repository LICENSE file
• Benchmarks: https://www.computesdk.com/benchmarks, with raw results under results/ in the benchmarks repo
• HN threads: https://news.ycombinator.com/item?id=45119931 and https://news.ycombinator.com/item?id=48385990