Anthropic announced, from London, on May 19 two new features for its agentic platform: self-hosted sandboxes and MCP tunnels. The JDN on site deciphers these announcements for you.
Anthropic wants to become the operating system for artificial intelligence agents. On the occasion of Code with Claude London, its first developer conference outside the United States on May 19, Claude’s publisher is announcing two building blocks that complete its Managed Agents platform: self-hosted sandboxes, in public beta, and MCP tunnels, in preview. A strategic announcement that sheds light on the positioning of the start-up.
Truly autonomous AI agents
Launched in April, Claude Managed Agents is Anthropic’s new agent platform. The principle is to prevent developers from reinventing the agentic loop, tool execution and runtime infrastructure. And thus facilitate the creation of a complete, truly autonomous agent. Where the classic API simply exposes the model, Managed Agents provides a turnkey framework. The developer creates an agent (template, system prompt, tools, MCP servers, skills), configures a cloud container environment, then launches sessions that run autonomously, sometimes for several hours, with bash, file operations, web search and MCP.
The runtime moved locally
Until now, the runtime, the execution environment where the agent launches its commands, reads and writes its files, ran exclusively on Anthropic servers. With self-hosted sandboxes, announced on May 19, the runtime can be operated locally. Only the agentic loop, which orchestrates the session, manages the context and error recovery, remains at Anthropic. The execution of the tools is carried out in the client’s infrastructure or at a managed sandbox provider. Four partners are supported at launch: Cloudflare, Daytona, Modal and Vercel.
A more secure MCP for businesses
The second brick, the MCP tunnels, completes the system. MCP, even though it has become the standard for connecting an agent to external tools or data, is still not very secure. Bringing an agent hosted at Anthropic into dialogue with an MCP server installed on a company’s internal network has until now required exposing this server to the Internet or setting up a VPN. An insecure or overly complex device. With MCP tunnels, it is now possible to deploy a gateway in your network with a direct outgoing connection to Claude Managed Agents. The agent can thus use the company data exposed via the tunnel. Launched in preview, the functionality is not only available on Claude Managed Agents but also in Anthropic’s historical API.
Weak signals that converge
Moving the runtime locally, securing the MCP for the company, choosing London for the first developer conference outside the United States: Anthropic is attacking head-on the European agentic AI market, historically more conservative on these subjects. The IT departments of the old continent make data sovereignty and control of the execution scope non-negotiable prerequisites. Today’s two announcements tick precisely these boxes. Furthermore, timing is not neutral. At a time when European regulators are closely scrutinizing the new capabilities of Claude Mythos and the new “frontier” AIs, a tightening of the AI Act can no longer be ruled out.
Showing now that its agents can operate within the client’s perimeter, with its own audit logs and its own network policies, is to get a head start on a debate that is coming quickly. That the demonstration was made from London and not from San Francisco is surely not a coincidence.
