Cloudflare & Anthropic launch Claude agents on sandboxes

Cloudflare and Anthropic have integrated Claude Managed Agents with Cloudflare Sandboxes, giving developers a Cloudflare-based environment in which to run Claude agents.

The core Claude agent loop remains on Anthropic's platform, while code execution, tool calls and network connections run on Cloudflare infrastructure. The setup is aimed at developers who want more control over where agent workloads run and how they connect to services.

In this model, a Claude agent session sends requests to a control plane built on Cloudflare Workers. The control plane assigns each session a sandboxed environment where the agent can execute code, use command-line tools and maintain state during periods of inactivity.

Developers can choose between a microVM-based sandbox and a lighter isolate-based option. The isolate route is designed for workloads that need to start quickly and run at higher volume, while microVMs remain available for agents that need a fuller Linux environment.

The companies are presenting the move as a way to separate model orchestration from the infrastructure used to carry out tasks. Anthropic has described that approach as "decoupling the brain from the hands."

The default deployment template includes traffic proxies, sandbox metrics and logs, browser tooling, private service access, email functions and support for custom tools. Users can also customise sandbox images, inspect running machines and connect logs to external monitoring services.

Infrastructure choice

The announcement reflects a broader shift in the AI tools market as companies seek more direct control over execution environments for agents. Managed agent products have typically bundled the model, orchestration layer and compute environment together. But businesses in regulated sectors, or those with specific internal systems, often want the execution layer closer to their own networks and policies.

The environment supports outbound proxies that can inject credentials outside the sandbox so agents do not directly handle secrets. That approach is intended to reduce the risk of data exfiltration when agents connect to external or internal systems.

Private connectivity is another part of the offering. The setup can connect agents to internal services through Cloudflare Mesh or Workers VPC, enabling access without exposing those services to the public internet.

Users can define egress rules for particular agents, tenants or services, including endpoint allowlists and policies for credential injection. Developers can also write custom proxy middleware to tailor how agents reach external systems.

Browser and email tools

The package also includes browser functions for web interaction, including browser search, browser execution, screenshots and content retrieval. These tools are designed to give developers a record of browser sessions, including session recording and workflows that allow human review.

Email support is built in through tools for sending, reading and listing messages. Agents can also be triggered through email and send mail using domains configured through Cloudflare's email service.

Other built-in functions include access to private services and image generation through Workers AI. Developers can also extend the setup with their own functions rather than build separate infrastructure for each tool.

Scale ambitions

Cloudflare used the launch to underline its view that agent software will require a cheaper, lighter execution layer than traditional virtual machines can provide on their own. The isolate-based option is intended for cases where organisations may need to run very large numbers of concurrent agents.

Agents built on Dynamic Workers using Codemode can execute arbitrary code inside a V8 isolate while still getting access to a file system. For workloads that require Linux-based developer tools or broader application building, Cloudflare Containers remain available.

That distinction points to a split emerging in the market between heavier agent environments suited to software development tasks and lighter environments tailored to high-volume automation. Cloudflare is aiming to address both within the same framework.

The integration also adds to Cloudflare's wider push into AI infrastructure through products including Sandboxes, Agents SDK, Browser Run, Dynamic Workers, Workers AI and networking tools for private connectivity. It has been expanding those services as it seeks to position its developer platform as a base for deploying and managing AI agents.

Developers can deploy the integration through a default template and then modify the repository to fit their own requirements. Each sandbox can be observed through the Cloudflare dashboard, with logs available for querying or export to services such as Datadog or Splunk.