Vercel Run SDK Isolates Untrusted Code From AI Agents.

Code generated by an AI agent can be useful right up until it gains access to production credentials, internal services or the wider network. Vercel is addressing that risk with Run SDK, a new package designed to execute untrusted JavaScript and TypeScript inside an isolated environment.

The company introduced Run SDK in an engineering post on August 25, 2026. It gives developers a controlled way to let agents write programs, coordinate tools and process results without handing that generated code the same privileges as the surrounding application.

That distinction matters as agent workflows become more capable. Calling JavaScript's eval function inside an application can expose everything available to that process, including secrets and internal systems. It also provides no durable mechanism for stopping execution when authentication or human approval is required.

Run SDK evaluates JavaScript, or TypeScript with its types removed, in a fresh QuickJS context running inside a worker thread. According to Vercel, that context has no direct route to Node.js or the network.

Instead, the application explicitly exposes selected operations through hostFunctions. These functions act as a narrow bridge between the isolated program and the host application. Agent-generated code can request an approved capability, such as interacting with an application service, while the host retains control over what is actually available.

This means web access and API interactions must be mediated rather than granted by default. The security boundary is consequently shaped by the host functions a developer chooses to expose. A narrowly defined function can limit an agent to a specific operation; an overly broad function could still create unnecessary access.

Developers can install the package with:

pnpm add run

The project is also available through Vercel's Run SDK GitHub repository.

Isolation is only part of the release. Run SDK can interrupt a program when it reaches a step requiring authentication or human-in-the-loop approval. Once a decision is supplied, execution can resume without repeating work that has already been completed.

That pause-and-resume model is useful for agents that mix low-risk processing with consequential actions. An agent might organize data autonomously but stop before accessing a protected account or triggering an external operation. Rather than forcing developers to rebuild the workflow around each checkpoint, Run SDK keeps the program's progress intact while the application handles the decision.

For small teams, this could reduce the amount of custom orchestration needed around agent-generated programs. It does not remove the need to define permissions carefully, validate inputs and decide which actions require review. What it provides is a more structured execution layer than running model output directly inside an application process.

The release also extends Vercel's growing set of infrastructure for agent-based applications. It follows work on secure connectivity and communication, including Vercel Connect adding Linq messaging for apps and agents. Run SDK focuses on a different problem: controlling the code an agent creates after a model has produced it.

The practical appeal is straightforward. Developers can give an agent room to write and execute multi-step logic, while keeping Node.js, network access and sensitive application capabilities outside the default execution environment. Access only crosses that boundary through functions the host application has deliberately provided.

What is Vercel Run SDK?

Vercel Run SDK is a package for executing untrusted JavaScript and type-stripped TypeScript generated by AI agents. It runs code in a fresh QuickJS context inside a worker thread and lets applications expose specific capabilities through controlled host functions.

Does Run SDK give agent-generated code direct network access?

No. Vercel says the execution context has no direct route to Node.js or the network. Applications can instead provide selected host functions that mediate approved operations, including interactions with external services.

When is Run SDK available?

Vercel announced Run SDK on August 25, 2026. Developers can install it using pnpm add run, and the source is available through the project's GitHub repository; the announcement does not specify separate pricing or plan requirements.