Anthropic Launches Self-Hosted Sandboxes and MCP Tunnels for Claude Managed Agents
Original: Anthropic Launches Self-Hosted Sandboxes and MCP Tunnels for Claude Managed Agents View original →
Overview
At its Code with Claude London event, Anthropic announced two new security-focused features for Claude Managed Agents: self-hosted sandboxes and MCP tunnels. Both are designed to let enterprises operate AI agents while keeping sensitive files, internal services, and proprietary data entirely within their own perimeter.
Self-Hosted Sandboxes
Available now in public beta, self-hosted sandboxes move tool execution from Anthropic's infrastructure to the customer's own environment. When Claude agents need to run code, browse files, or interact with services, those actions happen inside the company's infrastructure — files and repositories never leave the enterprise boundary. For organizations that prefer managed options, Anthropic supports Cloudflare, Daytona, Modal, and Vercel as certified sandbox providers.
MCP Tunnels
In research preview (access by request), MCP tunnels let agents connect to MCP servers on private networks without exposing those servers to the public internet. A lightweight gateway opens a single outbound, end-to-end encrypted connection — no inbound firewall rules or public endpoints needed. This enables agents to securely access internal databases, private APIs, and ticketing systems as tools.
What This Means for Enterprise AI
The announcement lowers the barrier for regulated industries — finance, healthcare, legal — to adopt Claude agents at scale. Agent orchestration still runs on Anthropic's servers, but actual tool execution happens entirely within the enterprise boundary. The features signal a broader push by AI companies to build the trust infrastructure needed to meet enterprise security and compliance standards.
Related Articles
Anthropic released ten ready-to-run agent templates for financial services including pitchbook creation, KYC screening, and month-end close. Claude now works directly in Excel, PowerPoint, Word, and Outlook.
Anthropic analyzed millions of real Claude interactions and found the 99.9th percentile session duration nearly doubled to 45+ minutes in 3 months, with software engineering accounting for nearly half of all agentic use.
Anthropic says Claude contains a J-space that resembles a global workspace for active, verbalizable thoughts. The lead tweet has more than 9.1 million views and points to audit use cases, including hidden goals in sabotage-trained models.