Independent developer Austin Wang has released CloudRouter, a CLI tool that allows AI coding agents like Claude Code and Codex to provision and manage their own cloud virtual machines and GPUs. The project, posted to Hacker News on February 13, 2026, addresses a fundamental constraint in agent-based development: resource isolation and scalability.
CloudRouter Solves Agent Resource Conflicts
When AI agents write and test code, they typically operate on the developer's local machine, sharing ports, RAM, and screen real estate. This creates conflicts when running multiple agents in parallel. While Docker provides some isolation, it still consumes local resources and lacks GPU access or browser capabilities agents need to verify their work.
CloudRouter provides a primitive that gives each agent its own dedicated environment. Agents can start VMs from local project directories, upload files, execute commands remotely, and tear down instances when finished. GPU requests are handled with simple flags.
Every VM Includes Development Tools and Remote Access
Each CloudRouter VM comes preconfigured with a VNC desktop, VS Code, and Jupyter Lab, all accessible through authenticated URLs. This allows developers to observe agents performing browser automation in real time through the VNC interface. The tool wraps Vercel Labs' agent-browser library for browser control.
GPU options include T4 (16GB VRAM), A100 (40GB VRAM), and H100 (80GB VRAM) instances. Commands follow a straightforward syntax:
cloudrouter start ./my-project— Launch a standard VMcloudrouter start --gpu B200 ./my-project— Launch with GPUcloudrouter ssh cr_abc123 "npm install && npm run dev"— Execute remote commands
Browser Automation Enables Visual Verification
CloudRouter's browser automation features allow agents to interact with web applications and capture snapshots:
cloudrouter browser open cr_abc123 "http://localhost:3000"— Open browser on VMcloudrouter browser snapshot -i cr_abc123— Extract interactive elementscloudrouter browser click cr_abc123 @e2— Trigger interactionscloudrouter browser screenshot cr_abc123 result.png— Capture visual state
The element snapshot feature returns labeled interactive elements (links, buttons, inputs) that agents can reference in subsequent commands.
Workflow Inversion Enables Parallel Agent Operations
Wang notes that CloudRouter inverts traditional cloud development workflows. Rather than starting from cloud-based agents and pulling work local for testing, CloudRouter keeps agents running locally while pushing their workloads to the cloud. "As I stopped watching agents work and worrying about local constraints, I started to run more tasks in parallel," Wang wrote in the Hacker News post.
The GPU provisioning capability eliminates manual setup for training and inference tasks. Agents can spin up GPU sandboxes, execute workloads, and clean up automatically. Early users have deployed multiple agents running experiments in parallel.
CloudRouter is open source and available in the manaflow-ai/manaflow monorepo on GitHub under packages/cloudrouter. A demonstration video is available on YouTube.
Key Takeaways
- CloudRouter enables AI coding agents to provision and manage their own cloud VMs and GPUs through simple CLI commands
- Each VM includes VNC desktop, VS Code, and Jupyter Lab with authenticated remote access for real-time monitoring
- Browser automation wraps Vercel Labs' agent-browser library, allowing agents to interact with web interfaces and capture screenshots
- GPU support includes T4, A100, and H100 instances that agents can request and tear down automatically
- The tool inverts traditional workflows by keeping agents local while pushing their workloads to isolated cloud environments for parallel execution