Anthropic: MCP Tunnels, Self-Hosted Sandboxes and Automatic File-Spill for Agents

Anthropic announced three new Claude API platform updates on May 19, 2026, targeting agent builders and enterprise teams deploying complex AI workflows.

What Are MCP Tunnels and Why Do They Matter?

MCP (Model Context Protocol) is an open protocol that defines how AI models communicate with external tools and data sources through a standardized interface. An MCP server is a component that implements this protocol and exposes specific functionality — such as access to a database, internal documents, or business systems.

Until now, connecting Claude agents to custom MCP servers required those servers to be publicly accessible on the internet. This posed a security risk for companies keeping sensitive internal tools behind corporate firewalls.

MCP Tunnels solves this problem: the feature, currently available as Research Preview, allows direct connections to MCP servers within private networks — without the need to expose ports or navigate complex VPN configurations. An agent can communicate with internal tools as if they were publicly available, while traffic remains within protected infrastructure.

Research Preview indicates early access to a feature that is functional but still collecting user feedback.

Who Controls Where Agents Execute Code?

Claude Managed Agents is Anthropic’s managed platform for running Claude as an autonomous agent with built-in tools, sandboxed execution and streaming responses. Until now, all agents executed tools exclusively on Anthropic’s infrastructure.

The new update introduces self-hosted sandboxes as a full alternative. A sandbox is an isolated execution environment where an agent can run code and tools without risk of unintended side effects on the rest of the system. Companies can now run that sandbox on their own servers — opening the door to stricter data residency requirements, lower latency for local resources, and full control over compute infrastructure costs.

This change is particularly relevant for financial services, healthcare and government, where regulatory frameworks often prohibit processing data on external infrastructure.

Automatic File-Spill: End of Manual Management for Large Outputs

The third update addresses a practical problem faced by teams building agents for document analysis, log processing or large dataset work. When a tool output (agent_toolset or MCP tool) exceeds 100K tokens, the system no longer truncates the content or returns an error.

Instead, Claude Managed Agents automatically saves the full output to a file inside the sandbox. The model receives a shortened preview with the exact file path and can read the full content in the next step — selectively if needed. This eliminates manual token window management when working with large tool results.

What Else Does This Release Include?

Alongside the three features above, Anthropic has added the ability to update MCP server and tool configurations within an active agent session — without restarting it. This accelerates iterative development and debugging of complex agent workflows.

All four updates are available through the existing managed-agents-2026-04-01 beta header, and detailed documentation for MCP Tunnels and self-hosted sandboxes has been published on the Anthropic Claude platform.