Vercel Drives bring persistent storage to Sandboxes in public beta

Persistent storage arrives for Vercel Sandboxes

Vercel has moved Drives for Vercel Sandbox into public beta, making the feature available to Hobby, Pro and Enterprise customers alike, according to the Vercel blog. A Drive is a persistent storage volume that mounts as a directory inside a sandbox. Because it is not tied to a single sandbox, the same Drive can be attached across separate runs and to different sandbox instances.

How mounting works

The workflow runs through Vercel's sandbox SDK. A developer creates or fetches a Drive by name, then references it in the mount configuration when starting a sandbox. From that point the Drive behaves like any other directory: applications read and write files through the sandbox filesystem at the mount path, and anything stored there survives after the sandbox stops.

Vercel frames the feature around agent workloads. Typical uses include keeping an agent's working files and on-disk context between sessions, and reusing datasets, models and dependency trees so each new sandbox does not have to rebuild them from scratch.

One writer, many readers

Concurrency follows a specific model. According to the Vercel blog, a Drive supports exactly one read-write mount at a time. However, once a Drive has been written to at least once, multiple sandboxes can read from it simultaneously by mounting point-in-time, read-only snapshots.

Each snapshot captures the Drive's state at the moment it is mounted. Writes that happen afterwards are not visible to existing snapshots, so consumers that need newer data must mount a fresh one. The design enables a fan-out pattern: one sandbox prepares or updates the data while several others — say, review or test sandboxes in an agent pipeline — read a consistent view in parallel.

Limits, regions and pricing

Several constraints apply. A single sandbox can mount up to four Drives at separate paths. Drives default to a maximum size of 1 TiB — 1 GiB on the Hobby plan — and can be configured up to 16 TiB, with higher limits available on request.

Drives exist in every Sandbox region, but they are region-bound: a Drive stays in the region where it was created, and any sandbox mounting it must run in that same region. Those sandboxes cannot use failover regions.

Pricing is usage-based, covering storage, reads and writes, with rates that vary by region. In the iad1 region, storage costs $0.05 per GB-month, reads $0.0015 per GB and writes $0.004 per GB. Hobby plans include an allowance of 15 GB of Drive storage plus 30 GB each of reads and writes per month. Vercel points to its Sandbox pricing page for the full regional breakdown.

Why it matters

Sandboxes are short-lived by design — files stored inside them disappear when the run ends — and that ephemerality has been a double-edged sword for AI agents. A fresh environment avoids state drift, but it also means an agent's work vanishes between sessions. Drives close that gap by giving sandboxes a durable, shared place to keep workspaces, models and data, cutting redundant setup work in pipelines that spin up many short-lived environments.

The single-writer, snapshot-reader model is a pragmatic fit for those pipelines, letting one sandbox produce data while many others consume a stable view of it. The trade-offs are real: region pinning removes failover flexibility, and metered reads and writes mean heavy data shuffling between sandboxes will show up on the bill. For teams building agent infrastructure on Vercel, the beta is a meaningful step toward making Sandboxes a credible home for stateful work.