Cloudrouter Opens Cloud Sandboxes and Browser Automation to AI Agents

Developers working with AI coding assistants like Anthropic's Claude Code now have a powerful new tool for cloud-based development. Manaflow AI's recently launched Cloudrouter provides a CLI interface that allows AI agents to spin up cloud sandboxes, execute commands, transfer files, and automate browser interactions—all without manual infrastructure management.

Bridging the Agent-Cloud Gap

Until now, AI coding assistants operated within constrained local environments, unable to interact with cloud resources or browsers. This limitation prevented agents from testing cloud deployments, running GPU-accelerated workloads, or automating web interactions. Cloudrouter addresses this by exposing cloud infrastructure through a unified command-line interface that AI agents can leverage as a skill.

The tool connects to multiple sandbox providers including E2B (default) and Modal (GPU support), provisioning environments ranging from 2 vCPU instances to multi-GPU configurations. Developers install it either as a standalone CLI via npm install -g @manaflow-ai/cloudrouter or directly integrate it into AI agents using npx skills add manaflow-ai/cloudrouter.

Core Capabilities

Instant Cloud Environments

• Spin up sandboxes from local directories or Git repos with automatic file syncing
• Specify GPU types (T4, A100, H100) and quantities (e.g., --gpu H100:2)
• Access via VS Code browser, VNC desktop, or terminal sessions
• Docker support with custom image overrides

Browser Automation Engine

• Built on Vercel's Agent Browser
• Full Chrome DevTools Protocol access for navigation, interaction, and scraping
• Element targeting through accessibility tree references (@e1) or semantic locators
• State persistence for authenticated sessions

File Operations

• Bidirectional sync between local and sandbox environments
• Watch mode for automatic re-uploading on file changes
• Exclusion patterns for large directories like node_modules

Technical Architecture

Cloudrouter acts as a middleware layer between AI agents and cloud providers. When an agent executes a command like cloudrouter start --gpu A100 .:

1. Authenticates via OAuth2 flow (cloudrouter login)
2. Provisions GPU instance on Modal or E2B
3. Syncs local directory to /home/user/workspace
4. Exposes access endpoints (VSCode, VNC, SSH)

The browser automation leverages Chrome instances running within sandboxes, with all commands executed via SSH tunnels. A notable implementation detail: element interaction requires sequential snapshot commands before and after DOM changes to maintain reference validity.

GPU Support Matrix

GPU	VRAM	Use Cases
T4	16GB	Small model inference
A100	40GB	7B-70B model training
H100	80GB	Research & fast training
H200	141GB	Max memory workloads

Security Design

• All development server ports (e.g., Vite's 5173) remain private
• Access only through authenticated VNC/Code sessions with token protection
• Explicit prohibition against exposing raw E2B port URLs
• Sandbox isolation with 10-instance concurrency limits

Practical Applications

1. AI Agent Enhancement: Claude Code can now prototype web apps end-to-end—spinning up environments, installing dependencies, and testing UIs via browser automation
2. ML Workflows: Researchers instantly provision multi-GPU instances for distributed training without cloud console access
3. Visual Testing: Automated screenshot comparisons across browser states
4. Education: Persistent sandboxes for workshop participants

Challenges and Considerations

The tool requires careful reference management during browser automation—stale element references after DOM changes necessitate frequent re-snapshotting. Additionally, npm permissions in new sandboxes require an initial chown command to avoid install failures. For teams needing >10 concurrent instances, Manaflow encourages direct contact for quota increases.

As AI agents evolve beyond code generation toward full-stack implementation, tools like Cloudrouter provide the missing infrastructure layer. By abstracting cloud complexity into CLI commands, it enables a new category of agent-driven development workflows while maintaining security boundaries. The MIT-licensed project is available on npm with source expected to follow.