Skip to main content

Run Claude Managed Agents with Vercel Sandbox

By Steven Van ·

Each session runs in an isolated Firecracker microVM with credential brokering and deny-by-default network egress.

Vercel now lets developers self-host the execution environment for Claude Managed Agents on Vercel Sandbox. Anthropic still runs the model, harness, tools and session state, but an agent's tool calls can now execute on a developer's own Vercel infrastructure, with access to private APIs, internal services and customer data.

Each agent session runs in its own isolated Firecracker microVM, spawned by a Vercel function control plane. What's included:

  • A managed agent loop from Anthropic: model, harness, tools and session state
  • A Vercel function control plane that spawns a sandbox per session
  • Firecracker microVM isolation per session, with millisecond startup
  • Credential brokering at the firewall, so secrets never enter the sandbox
  • Deny-by-default egress with a domain allowlist
  • Low-latency connectivity to a developer's private network and cloud workloads
Vercel
Vercel
The platform for frontend developers — deploy, preview, and scale web apps and AI agents with zero config.
View Vercel →

Read the original announcement →

Read Run Claude Managed Agents with Vercel Sandbox on Creators Toolbox