Anthropic Ships Enterprise Agent Infrastructure: Self-Hosted Sandboxes, MCP Tunnels, and a 30,000-Person PwC Deployment

In one week, Anthropic launched self-hosted agent sandboxes, private network tunnels, acquired SDK maker Stainless, and began deploying Claude across PwC’s global workforce with 70% delivery time cuts.
artificial-intelligence
software-engineering
Author

Kabui, Charles

Published

2026-05-26

Keywords

anthropic-enterprise, claude-managed-agents, self-hosted-sandboxes, mcp-tunnels, stainless-acquisition

Download as PDF

In one week, Anthropic shipped the infrastructure to turn Claude from a model API into a full enterprise agent platform. On May 19, self-hosted sandboxes entered public beta for Claude Managed Agents: tool execution runs on your infrastructure through providers like Cloudflare, Daytona, Modal, or Vercel, while Anthropic handles the orchestration loop. The same update introduced MCP tunnels in research preview, letting agents reach private MCP servers inside corporate networks through a single encrypted outbound connection, no inbound firewall rules needed. A day earlier, Anthropic acquired Stainless, the company that has generated every official Anthropic SDK since launch across TypeScript, Python, Go, Java, and Kotlin. Stainless also builds the CLIs and MCP server connectors that hundreds of companies rely on.

Claude Code and Claude Cowork are rolling out to PwC’s global workforce of hundreds of thousands, with 30,000 U.S. staff being trained and certified. Insurance underwriting that took 10 weeks now takes 10 days. A COBOL mainframe migration four times larger than scoped is tracking on time and under budget. PwC is launching an entire finance business unit, Office of the CFO, built on Claude. On the legal side, adoption doubled to 87% among general counsel in 2026, and Claude now offers 12 pre-built practice-area plugins for workflows like contract review, litigation prep, and regulatory monitoring.

Anthropic now controls the model, the SDK layer (via Stainless), the agent orchestration (Managed Agents), and the connectivity standard (MCP). Self-hosted sandboxes solve the code execution trust problem. MCP tunnels solve the network access problem. The list of enterprise objections to AI agent deployment is getting short.

Read More: AI This Week: OpenAI Disproves an 80-Year Math Conjecture, Karpathy Joins Anthropic, Google Drops Gemini 3.5, OpenAI Files for IPO

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {Anthropic {Ships} {Enterprise} {Agent} {Infrastructure:}
    {Self-Hosted} {Sandboxes,} {MCP} {Tunnels,} and a {30,000-Person}
    {PwC} {Deployment}},
  date = {2026-05-26},
  url = {https://toknow.ai/posts/anthropic-enterprise-agent-infrastructure-sandboxes-mcp-tunnels-pwc/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “Anthropic Ships Enterprise Agent Infrastructure: Self-Hosted Sandboxes, MCP Tunnels, and a 30,000-Person PwC Deployment.” https://toknow.ai/posts/anthropic-enterprise-agent-infrastructure-sandboxes-mcp-tunnels-pwc/.