Anthropic has rolled out significant enhancements to its Claude Managed Agents, introducing self-hosted sandboxes and MCP tunnels. These new features aim to give enterprises increased autonomy over their AI agents' operational environment while retaining key management components within Anthropic’s infrastructure.
The self-hosted sandboxes enable companies to execute AI tools within their own systems, ensuring that files and repositories remain secure and internal. This capability allows organizations to uphold their existing network policies, audit logs, and security measures. Companies can also customize their computing resources by selecting their preferred CPU, memory, and runtime images. For those unable to establish their own infrastructure, Anthropic has partnered with managed service providers such as Cloudflare, Daytona, Modal, and Vercel, offering an alternative path for integration.
In tandem, Anthropic introduced MCP (Model Context Protocol) tunnels, which create secure connections between agents and MCP servers over a private network. This feature bypasses the public internet, using a lightweight gateway for a single, end-to-end encrypted outbound connection. The design eliminates the need for inbound firewall rules or public endpoints, allowing agents to access internal databases, private APIs, and ticketing systems as tools.
Despite these advancements, Anthropic maintains control over the core functionalities of the agents. The orchestration of agent context management, error handling, and overall execution loop remains on Anthropic’s servers. Consequently, businesses seeking to fully manage model execution will encounter limitations within this framework.
Both features are currently in early stages: the self-hosted sandboxes are accessible as a public beta, while MCP tunnels are available only as a research preview, requiring companies to apply for access. This indicates that while Anthropic is enhancing enterprise capabilities, full autonomy over the AI agents is not yet achievable.
As organizations increasingly integrate AI into their operations, the demand for secure and controlled environments grows. Anthropic's latest offerings reflect an understanding of these needs, positioning the company as a key player in the AI crypto sector. However, their approach to retaining control over the agent loop raises questions about the future of enterprise autonomy in AI deployment. The balance between security and operational flexibility will likely be critical as more companies explore the potential of AI-managed agents in their workflows.
