AI code-execution sandbox

Vercel Sandbox

Vercel Sandbox is a Firecracker-based, JavaScript/TypeScript-native sandbox that is persistent by default, billed on active CPU plus a per-creation fee.

Last verified 2026-07-26

Vercel Sandbox brings secure code execution into the Vercel platform for AI apps. Firecracker isolation, JS/TS-native, persistent by default, single-region. Session caps are 45 minutes on Hobby and 24 hours on Pro/Enterprise.

Key specs

IsolationFirecracker microVM
Cold start400ms
Price / vCPU-hr$0.128
Pricing modelActive CPU + per-creation fee
GPUNo
Session limit45min (Hobby) / 24h (Pro)
PersistencePersistent by default
LanguagesJavaScript/TypeScript (native), Python
MCP supportn/a
IntegrationsOpenAI Agents SDK, Vercel AI SDK
DeploymentManaged (Vercel platform)
Free tierIncluded allowance on Hobby
Active CPU$0.128/vCPU-hr
Creations$0.60 / 1M

Strengths

  • Tight Vercel/AI SDK integration
  • Firecracker isolation
  • Persistent by default

Trade-offs

  • Single region
  • Creation fee ($0.60/1M)
  • Short Hobby session cap

Best for: Teams already on Vercel building JS/TS AI apps.

How Vercel Sandbox works

Vercel Sandbox runs on Firecracker microVMs, so each sandbox gets a real kernel boundary, and it is built for JavaScript and TypeScript first, with Python also supported. Sandboxes are persistent by default rather than torn down after a run. Cold start is around 400ms. It runs in a single region, and session length is capped at 45 minutes on Hobby and 24 hours on Pro and Enterprise. The natural home for it is a team already on Vercel building AI apps, where it slots into the AI SDK and the platform they already run.

Pricing in practice & watch-outs

Billing has two parts: active CPU at about $0.128 per vCPU-hour, plus a per-creation fee of $0.60 per million sandboxes created. If your pattern is many short-lived sandboxes, that creation fee is worth modelling, since it is charged separately from compute. Watch the Hobby session cap of 45 minutes, which is tight for longer agent runs; you need Pro for the 24-hour limit. It is also single-region, so latency and data residency are fixed to one location. The value case is strongest if you already pay for Vercel.

Head-to-head

Alternatives to Vercel Sandbox

See the full Vercel Sandbox alternatives comparison →

Sources: Vercel Sandbox docs ↗

Is Vercel Sandbox right for your workload?

Tell us your use case and we'll confirm the fit, flag anything to watch for, and get you real pricing from the Vercel Sandbox team. No spam.