Anthropic Strengthens Digital Sovereignty from London

Anthropic Unveils Innovations in London

On May 19, during an event in London, Anthropic unveiled two new features for its intelligent agent platform: self-hosted sandboxes and MCP tunnels. These announcements mark a significant milestone for the startup, which aims to establish itself as a key player in the field of operating systems for artificial intelligence agents.

At the "Code with Claude London" conference, the first of its kind organized by Anthropic outside the United States, the company presented these new technological components that enhance its Managed Agents platform. The self-hosted sandboxes are currently in public beta, while the MCP tunnels are available in preview. These strategic developments underscore Anthropic's commitment to strengthening its market position.

Towards Total Autonomy for AI Agents

The Claude Managed Agents platform, launched in April, aims to simplify developers' work by eliminating the need to recreate the fundamental elements required to create intelligent agents. This platform offers a comprehensive framework that allows for the design of truly autonomous agents. Unlike a traditional API that merely provides the model, Managed Agents offers a suite of integrated tools. Developers can create an agent by configuring a model, a system prompt, tools, MCP servers, and specific skills, all within a cloud container environment. Sessions can operate autonomously for several hours, utilizing tools such as bash, file operations, web searches, and MCP connections.

Local Execution of the Runtime

Until now, the execution environment, or runtime, where the agent executes its commands, was exclusively hosted on Anthropic's servers. With the introduction of self-hosted sandboxes, announced on May 19, companies can now operate the runtime locally. Only the agent loop, which orchestrates the session and manages errors, remains under Anthropic's management. The execution of tools can occur on the client's infrastructure or via a managed sandbox provider. For the launch, four partners are supported: Cloudflare, Daytona, Modal, and Vercel.

Enhanced Security with MCP Tunnels

The second innovation, MCP tunnels, aims to improve the security of connections. Although the MCP has become a standard for connecting an agent to external tools or data, it still presents security vulnerabilities. Until now, for an agent hosted at Anthropic to access an MCP server on a company's internal network, it was necessary to expose that server to the Internet or use a VPN, solutions that are often insecure or complex. The MCP tunnels now allow for the deployment of a gateway that establishes a direct outgoing connection to Claude Managed Agents, facilitating secure access to the company's internal data. This feature is available not only on Claude Managed Agents but also through Anthropic's legacy API.

A Strategy Focused on Europe

By choosing to move the runtime locally and secure the MCP for businesses, Anthropic is clearly targeting the European market, known for its commitment to data sovereignty and infrastructure control. These new features meet the expectations of European CIOs, who consider these aspects essential. Furthermore, the choice of London for this announcement is significant, as European regulators closely examine the capabilities of new AIs, such as Claude Mythos.

This strategy allows Anthropic to position itself ahead of upcoming regulatory discussions, particularly around the AI Act. By demonstrating that its agents can operate in the client's environment, with their own security policies and audit logs, Anthropic anticipates future requirements. The fact that this demonstration takes place in London, rather than San Francisco, highlights the importance of the European market for the company.