Anthropic took the stage at the Code w/ Claude 2026 conference in London on Tuesday to roll out updates to its Claude Managed Agents platform. The new features—self-hosted sandboxes and MCP tunnels—are aimed at giving developers more control over how they deploy and secure AI agents in production.
What the updates include
Self-hosted sandboxes let teams run Claude agents in environments they manage themselves, rather than relying solely on Anthropic’s cloud infrastructure. That means data stays inside the company’s own network, a selling point for industries with strict compliance requirements like finance, healthcare, or defense.
MCP tunnels, short for “Model Context Protocol” tunnels, provide a secure way to connect Claude agents to internal APIs and databases without exposing those systems to the public internet. The tunnels handle authentication and encryption under the hood, reducing the setup work developers usually face when wiring an agent to existing tools.
Both features were demonstrated during a live coding session at the conference, where an Anthropic engineer showed how a Claude agent could query a private sales database through an MCP tunnel while running inside a self-hosted sandbox.
Why enterprises have been asking for this
Since launching Managed Agents earlier this year, Anthropic has heard from large customers that they want more granular control over the agent’s runtime environment. The standard cloud-hosted version works fine for prototyping, but companies rolling out agents at scale often need to keep sensitive data behind their own firewalls.
Self-hosted sandboxes address that directly. Developers can spin up a sandbox in their own Kubernetes cluster or on a dedicated VM, then deploy Claude agents there. The sandbox isolates each agent from others on the same machine, preventing accidental data leaks between different projects or client accounts.
The MCP tunnels solve a different pain point. Previously, connecting an agent to a private API required setting up a VPN, configuring firewall rules, and managing API keys—a lot of overhead for a team that just wants to test a new agent workflow. With tunnels, the agent opens a secure outbound connection to Anthropic’s relay service, which then proxies requests to the internal endpoint. No inbound ports need to be opened.
What’s still unclear
Anthropic did not announce pricing for the self-hosted option during the London event. The company currently charges per agent-hour for the cloud-hosted version, and a spokesperson said only that “pricing for self-hosted will be competitive.” No release date beyond “later this year” was given for general availability.
Developers at the conference had mixed reactions. Some praised the tunnel feature as a long-overdue simplification, while others questioned whether running a self-hosted sandbox would require more ops staff than the typical AI team has on hand.
The conference continues through Thursday, with additional sessions on agent safety and fine-tuning scheduled. Anthropic has not said whether more Managed Agent features will be revealed before the closing keynote.
