Anthropic Claude Platform: Managed Agents Get Self-Hosted Sandboxes, Live MCP Config Updates, and Large Output Spilling

What's new

• Self-hosted sandbox: tool execution can now run in the developer's own infrastructure instead of Anthropic-managed sandbox
• Large output spilling: agent_toolset and MCP tool outputs exceeding 100K tokens are automatically spilled to a file in the sandbox; the model receives a truncated preview with the file path
• Live MCP config update: MCP server and tool configurations can be updated during an active session without restart
• MCP Tunnels (Research Preview): encrypted tunnel for connecting MCP servers in private networks without requiring a public IP
• Claude Managed Agents on AWS: webhooks and multi-agent orchestration now available on the Claude Platform on AWS
• Web search + SEC data: web search tool returns richer SEC filing data for financial research agents

Why it matters

Self-hosted sandboxes and MCP Tunnels are critical for enterprise deployments where security requirements prohibit routing tools through external servers. Large output spilling solves a practical context window problem in long agentic workflows.

How to try it

Available via the Claude Developer Platform (platform.claude.com). MCP Tunnels are in Research Preview — registration required.