Self-Hosted Sandboxes is a public-beta feature of 📝Claude Managed Agents that moves tool execution off Anthropic's managed cloud containers and onto a customer-controlled sandbox runtime.
By default, Managed Agent sessions execute tool calls — bash commands, file reads, code runs — inside Anthropic-operated cloud containers. With Self-Hosted Sandboxes, the customer brings their own sandbox client or selects from supported partners: Cloudflare, Daytona, Modal, or Vercel. The orchestration layer — the Agent Loop, context management, and event streaming — continues running on Anthropic's infrastructure. Only the execution surface moves.
The architecture works because of Managed Agents' brain-hands decoupling. The harness running the 📝Claude Agent Loop treats each sandbox as a replaceable tool with a uniform execute(name, input) → string interface, so swapping Anthropic-managed containers for a self-hosted runtime is transparent to the loop. Customers gain control over the execution environment, data residency, network egress policies, and audit logging — at the cost of operating the sandbox infrastructure themselves. Pricing is unaffected on Anthropic's side: standard token rates and the $0.08 per session-hour Managed Agents fee still apply.
Two notable beta limitations: Self-Hosted Sandboxes is not yet available on Claude Platform on AWS, and the Memory research-preview feature is not yet supported in self-hosted sessions. Announced at Code with Claude London in May 2026 alongside MCP Tunnels.