AI HOT (Curated Pool)May 19, 2026, 17:09 UTC+8

Claude Managed Agents Add Self-Hosted Sandboxes and MCP Tunnels

Claude智能体托管平台新增自托管沙箱与MCP隧道功能

Anthropic added two updates to the Claude managed agents platform: self-hosted sandboxes are in public beta, and MCP tunnels are in research preview for private network database and API access.

Why it matters: HKR-H/K/R all pass: this is an official Anthropic Claude agent-platform update with two concrete mechanisms. It is below model-release weight, but strong enough for featured agent-infra coverage.

Read the original ↗Export Markdown

Take

Anthropic is patching the enterprise-agent boundary, not showing new model intelligence. Self-hosted sandboxes are in public beta; MCP tunnels are only in research preview. One controls code execution. The other reaches private databases and APIs. Claude Managed Agents cannot enter serious workflows until teams know where agents run, what they can touch, and who owns the blast radius.

The sequencing is pretty clear. Anthropic already has Claude Code, Slack, Microsoft 365, and Skills as surfaces. Now it is filling in the runtime and network layer underneath them. OpenAI has been chasing enterprise connectors too, but Anthropic is packaging the agent runtime itself as the product. I don’t buy the production-readiness vibe yet: the post gives no permission model, audit granularity, or network isolation detail. Without those, an MCP tunnel is just a polished hole into the private network.

Source: AI HOT (Curated Pool) · claude.com
Published: May 19, 2026, 17:09 UTC+8 · May 19, 2026