Anthropic managed agents now include two enterprise features, MCP Tunnels and Self hosted Sandboxes, the company announced, a move vendors and buyers say pushes AI agents closer to practical enterprise deployment.
Anthropic managed agents address deployment challenges
AI agents differ from chatbots because they can proactively execute multi step workflows, fetch internal documents, call system APIs, and compile reports.
Enterprises have faced two persistent barriers to deploying agents: agents need access to internal network data, and connecting internal systems to external cloud services raises security and compliance concerns.
What MCP Tunnels does
Anthropic said MCP Tunnels builds on the Model Context Protocol, or MCP, which standardizes how agents connect to external tools and data sources.
MCP Tunnels creates a secure connection tunnel that lets Claude agents access corporate networks without requiring firms to expose internal systems directly to the public internet.
Research preview and limits
Anthropic released MCP Tunnels as a research preview, and the company said documentation, stability, and formal technical support are still being expanded.
Two people familiar with enterprise evaluations, who requested anonymity because they were not authorized to speak publicly, said companies should avoid using the preview for mission critical workflows until the feature is production hardened.
How Self hosted Sandboxes work
Anthropic described Self hosted Sandboxes as a way for customers to deploy agent tool execution and code runtime environments on their own infrastructure rather than on Anthropic cloud servers.
That design keeps corporate data inside the company environment during processing, which Anthropic said can help address data sovereignty and compliance concerns for regulated industries.
Shared responsibility for security settings
Anthropic emphasized that identity management, network segmentation, and tool permissioning remain the responsibility of the customer, and that the Managed Agents platform will not automatically configure those controls.
Two cybersecurity consultants who asked not to be named, because their firms consult with enterprise buyers, said the change provides important infrastructure primitives but does not remove the need for careful security governance.
Market positioning and competition
The update arrives as competition among enterprise AI platforms intensifies, with Microsoft Copilot and other large vendors expanding agent and assistant offerings.
Anthropic is positioning its approach around an open protocol combined with options to run components on customer infrastructure, effectively returning more control to enterprises compared with platforms that emphasize deep cloud integration, company materials show.
Analysts and buyers watching deployments said the combination of MCP Tunnels and Self hosted Sandboxes may accelerate trials for security conscious customers, but broader, stable rollout will depend on Anthropic completing productization and support for large scale, regulated workloads.
Anthropic provided the feature details in a company blog post and documentation pages, and did not announce a firm general availability date for MCP Tunnels.
