Anthropic Enhances Claude with Private Hosting and Secure Tunnels

Anthropic, a major player in the field of artificial intelligence, has recently expanded the capabilities of its Claude Managed Agents by introducing two new features: self-hosted environments and MCP tunnels. These additions aim to provide businesses with greater flexibility and security in the use of their AI agents.

Self-Hosted Environments

The self-hosted environments feature allows companies to run their AI agent tools on their own infrastructure. This means that files and repositories remain confined to the company's environment, thereby preserving data privacy and security. Businesses have the freedom to choose their own hardware configurations, such as CPU, memory, and runtime image. For those who prefer not to manage their own infrastructure, options from managed providers like Cloudflare, Daytona, Modal, or Vercel are available.

MCP Tunnels

MCP tunnels, or Model Context Protocol, represent the second new feature. This functionality enables agents to connect to private servers via a secure network, without exposure to the public Internet. A lightweight gateway ensures a single outgoing connection that is fully encrypted, eliminating the need for incoming firewall rules or public endpoints. This allows agents to access internal resources such as databases, private APIs, or ticketing systems.

Control and Limitations

Despite these advancements, Anthropic retains control over the orchestration of the agents, which still occurs on its own servers. Context management, error handling, and the agent loop remain under Anthropic's responsibility, meaning that a fully on-premises deployment of the agents is not feasible.

Availability and Access

Currently, self-hosted environments are available in public beta, allowing companies to test this feature. In contrast, MCP tunnels are still in research preview, and interested companies must request access to experiment with them.