Anthropic Adds Enterprise Control to Claude Managed Agents with Self-hosted Sandboxes and MCP Tunnels
anthropicclaudeenterprise-aiclaude-codeself-hostingsecurity
trending_upTrend: anthropic

Anthropic Adds Enterprise Control to Claude Managed Agents with Self-hosted Sandboxes and MCP Tunnels

calendar_month May 19, 2026 update Updated: May 22, 2026

🔄 Update — 22. May 2026: Anthropic and Atlassian Securely Automate DevOps Workflows

The partnership between Anthropic and Atlassian brings Claude Code directly into Bitbucket Agentic Pipelines. This enables secure automation of DevOps tasks, while the new self-hosted sandboxes ensure that sensitive corporate data never leaves the internal perimeter.

Was ist neu? / What’s new?

  • Bitbucket Integration: Claude Code can now be used directly within Bitbucket pipelines to create PRs, fix tests, and update documentation.
  • Security Guarantee: By moving tool execution to environments controlled by the company, the risk of credential leaks when using APIs is minimized.
  • MCP Tunnels Research Preview: Anthropic has released MCP Tunnels as a research preview to unify secure access to internal data sources.

Warum es den Artikel ergänzt / Why this adds to the article

This integration demonstrates the practical application of previously announced security features. It is no longer just about theoretical control, but about the productive use of agents in critical software development processes.

🔄 Update — 22. May 2026: Claude Code Evolves Toward Autonomous, Long-Running Workflows

Claude Code is showing a clear shift toward more autonomous behavior through the introduction of managed agents, sandboxing, and explicit goal-setting. Community experiments on platforms like Reddit and YouTube already demonstrate its use for complex, long-running coding tasks that go far beyond simple chat-based pair programming.

Was ist neu? / What’s new?

  • Managed Agents & Sandboxing: Official documentation introduces concepts for isolated execution environments (sandboxing) and goal-oriented agent control (goal).
  • Long-running Workflows: Developers are increasingly using Claude Code for proactive systems that work independently on repositories for hours or days.
  • Broad Signal Coverage: The trend is supported by official Anthropic documentation as well as technical coverage (InfoQ) and practical community use cases.

Warum es den Artikel ergänzt / Why this adds to the article

This development reinforces the thesis that Claude Managed Agents are taking control not only in enterprise workflows but also directly in software development. The line between assistant and autonomous agent is becoming increasingly blurred.

🔄 Update — 20. May 2026: Anthropic Transforms Claude into an Enterprise Agent Platform

Anthropic is evolving Claude from a pure AI model into a comprehensive platform for enterprise agents. By integrating it into enterprise subscriptions and collaborating with partners like KPMG, Claude is becoming more deeply embedded in operational workflows.

Was ist neu? / What’s new?

  • Claude Code Integration: Access to Claude Code is now officially formalized within Team and Enterprise plans, enabling professional management of developer workflows.
  • Enterprise Deployment at KPMG: KPMG is integrating Claude across its core business processes, underscoring the practical maturity of the agent infrastructure.
  • Security Focus: New privacy and security features reinforce Claude’s positioning as a controllable platform for sensitive corporate data.

Warum es den Artikel ergänzt / Why this adds to the article

These developments confirm the trend toward hybrid enterprise architectures described in the article and show that Anthropic is now heavily focusing on commercial operationalization.

Zusammenfassung / Summary

Anthropic has introduced two major security and privacy features for Claude Managed Agents: Self-hosted Sandboxes and MCP Tunnels. These features allow organizations to run AI agents within their own infrastructure and securely connect them to private, internal systems without compromising security standards.

Was ist passiert? / What happened?

With these new features, Anthropic addresses the biggest hurdles for enterprise AI agent deployment: security and data privacy.

  • Self-hosted Sandboxes: Code execution and tool interaction no longer happen in Anthropic’s cloud but within the customer’s own infrastructure or through partners like Vercel, Modal, or Cloudflare.
  • MCP Tunnels: A secure, outbound-only tunnel allows Claude to access internal databases and APIs without the need for complex inbound firewall rules.

Warum es wichtig ist / Why it matters

AI agents often fail security reviews because executing code in third-party environments poses high risks. By moving execution into the company’s own security perimeter, enterprises can apply existing audit logs, network policies, and zero-trust controls. This significantly accelerates the production deployment of AI workflows.

Beweise / Evidence

Anthropic officially announced these features in their documentation and across various tech platforms. Partners like Vercel and Cloudflare already support the integration. Self-hosted sandboxes are in Public Beta, while MCP tunnels are available as a Research Preview.

Analyse / Analysis

This move marks a shift from pure cloud models toward hybrid enterprise architectures. While the “intelligence” (orchestration) remains with Anthropic, the “action” (execution) becomes sovereign. This is a strategic move to win over enterprise customers who have been hesitant due to compliance concerns. However, a dependency on Anthropic remains as the model orchestration itself is not local.

Praktische Erkenntnisse / Practical Takeaways

  • Migration: Switching from Anthropic-hosted to self-hosted sandboxes typically requires only configuration changes.
  • Infrastructure Partners: Organizations can choose between specialized partners for web workloads (Vercel) or GPU-intensive tasks (Modal).
  • Security: Security teams can now define detailed network access policies for agents.

Offene Fragen / Open Questions

  • Latency: How much does the round-trip between Anthropic’s cloud and local execution affect response times?
  • Air-Gapped Systems: This solution is still insufficient for fully isolated environments as orchestration remains cloud-based.
  • Feature Gaps: Features like “built-in memory” are not yet supported in self-hosted environments.

Quellen / Sources

  1. Anthropic Introduces MCP Tunnels for Private Agent Access to Private Systems
  2. Self-hosted sandboxes - Claude API Docs
  3. Anthropic enhancements for Claude Managed Agents
  4. Anthropic adds self-hosted sandboxes and MCP tunnels
  5. Anthropic debuts MCP tunnels and self-hosted sandboxes
  6. Claude Code just got better | Agent View - YouTube
  7. Built a long-term autonomous agent system with Claude Code - Reddit
  8. Code with Claude - InfoQ
  9. Sandboxing - Claude Code Docs
  10. Goal - Claude Code Docs
  11. Build a proactive agent workflow with Claude Code - YouTube
  12. Claude agents can finally connect to enterprise APIs without leaking credentials
  13. New in Claude Managed Agents: self-hosted sandboxes and MCP tunnels
  14. Agentic Pipelines now supports Claude Code - Inside Atlassian