CloudRouter: Giving AI Coding Agents Their Own Cloud Infrastructure
The missing piece for autonomous software development might be letting agents spin up their own servers.
The Core Insight
Here’s the limitation nobody talks about enough: AI coding agents like Claude Code and Codex are remarkably capable at writing and analyzing code, but they’re essentially trapped in your local environment. Need to test a GPU-intensive ML pipeline? Hope your laptop has one. Want to spin up a clean environment to verify a deployment? You’re doing that manually.
CloudRouter changes this equation entirely. It’s a skill that lets coding agents provision their own cloud sandboxes — complete with GPU support, browser automation, and file synchronization — all through simple CLI commands. The agent doesn’t ask you to create a VM; it just does it.
cloudrouter start --gpu H100 .
That’s an AI coding agent provisioning itself a server with an 80GB VRAM H100 GPU. From your codebase. Without human intervention.
Why This Matters
Agents need execution environments, not just IDEs. The current wave of coding assistants is bottlenecked by their operating context. They can write perfect Kubernetes manifests but can’t actually deploy them. They can architect ML training pipelines but can’t run them. CloudRouter removes this constraint.
The GPU story is significant. The supported range spans from T4 (16GB, self-serve) through H200 (141GB) and B200 (192GB, “latest gen, frontier models”). Multi-GPU configurations are supported via simple syntax like --gpu H100:2. This isn’t toy infrastructure — it’s production-grade compute access.
Browser automation is built in. Every sandbox includes Chrome CDP integration. The agent can navigate websites, interact with elements via accessibility tree refs, take screenshots, and scrape data. This enables end-to-end testing workflows that weren’t previously possible for autonomous agents.
The skill paradigm is maturing. CloudRouter installs as a “skill” for Claude Code and other agents — a standardized way to extend capabilities. The documentation includes a complete skill reference that agents can ingest directly. We’re seeing the emergence of a plugin ecosystem for AI coding tools.
Key Takeaways
Multiple provider support. Currently E2B (default, Docker-focused) and Modal (GPU support), with Vercel, Daytona, Morph, and Freestyle on the roadmap. Provider abstraction means agents don’t need to learn multiple cloud APIs.
Rich access methods. VS Code in browser, Jupyter Lab, VNC desktop, interactive terminal, or one-off SSH commands. The agent picks what fits the task.
File sync is first-class. Upload, download, watch mode with auto-sync, exclude patterns, dry-run previews. Agents can maintain synchronization between local and remote environments programmatically.
Open source, MIT licensed. Built in Go, distributed as npm packages for macOS, Linux, and Windows. This isn’t a locked-down SaaS play — the code is inspectable.
Looking Ahead
The implications here extend beyond convenience. When AI agents can provision their own infrastructure, the definition of “autonomous software development” expands dramatically. An agent tasked with “optimize this ML model” could potentially:
- Spin up GPU instances
- Run training experiments in parallel
- Evaluate results
- Iterate on architecture
- Deploy the best performer
- Tear down temporary infrastructure
All without human involvement beyond the initial prompt.
We’re moving from “AI writes code that humans deploy” toward “AI writes, tests, and deploys code in environments it provisions itself.” The trust and supervision challenges this creates are substantial, but so is the productivity potential.
For teams building agentic development workflows, CloudRouter represents the infrastructure layer that makes ambitious automation actually feasible. The question isn’t whether agents will get cloud access — it’s how we’ll govern what they do with it.
Based on analysis of CloudRouter documentation by Manaflow AI (Feb 2026)