OpenClaw Machines: Scaling Enterprise AI Agents with Bare Metal

The Enterprise Challenge of AI Agent Infrastructure

As AI agents become increasingly sophisticated, the infrastructure required to run them securely, cost-effectively, and at scale presents a significant challenge for enterprises. Traditional approaches often fall short: running agents locally lacks isolation and scalability, Virtual Private Servers (VPS) offer limited kernel-level isolation and rising costs with scale, and managed services, while convenient, can lead to high per-agent costs and concerns over data sovereignty.

This is where OpenClaw Machines steps in. It's an open-source platform designed to address these gaps by enabling developers and organizations to self-host their OpenClaw agents on their own hardware, providing robust security through hardware isolation, predictable costs, and full control over data.

Introducing OpenClaw Machines: Your Private Cloud for AI Agents

OpenClaw Machines transforms your own Linux servers into a pool of secure, on-demand AI sandboxes. The core idea is simple yet powerful: each OpenClaw agent runs within its own dedicated Firecracker microVM. This provides true hardware isolation, leveraging the KVM hypervisor, ensuring that untrusted or agent-generated code is safely contained within its own guest kernel and memory space.

The platform consists of a control plane that orchestrates your hosts and manages the lifecycle of these microVMs, combined with a Cloudflare-powered data plane that acts as the secure front door. This data plane provides each machine with its own subdomain behind edge authentication, accessible via a Cloudflare Tunnel that terminates directly inside the VM, eliminating host port exposure for user traffic.

Architectural Deep Dive: Components and Flow

At its heart, OpenClaw Machines is a five-layer stack built on Apache-2.0 licensed components:

• Control Plane: Written in Go and backed by Postgres, this is the brain of the operation. It manages accounts, machines, and hosts, handles placement decisions, machine lifecycle (booting, stopping, reaping), host enrollment, backups, and durable workflows. While it manages agents, it still requires private or firewall-restricted access to each agent's authenticated control API on port 9090.

• Host Agent (ocm-agent): Deployed on your enrolled Linux servers, this agent supervises and manages the Firecracker microVMs. It handles tasks like booting and stopping VMs, managing bridge/TAP networking for connectivity, and staging the root file systems for each agent.

• MicroVMs and Runtimes:

  • Machine VMs: These are the isolated Firecracker microVMs where the OpenClaw agent resides. Inside, you'll find an auth proxy, a web-chat gateway for interaction, and a live terminal for direct access.
  • Browser VMs: For tasks requiring web automation, separate Firecracker microVMs run headful Chromium instances. These are driven by the main agent over the Chrome DevTools Protocol (CDP) and offer a watchable live view.

• LLM Proxy (LiteLLM): A per-host component, this proxy centralizes management of Large Language Model (LLM) API keys. It supports bringing your own keys (BYO-key), tracks per-machine usage across various providers, and crucially, allows routing traffic to open-source models running on your own GPU hardware, significantly reducing token spend.

• Workspace Integrations (Native MCP): This powerful feature allows you to connect external tools like GitHub, Google Workspace, OpenAPI endpoints, GraphQL APIs, or other remote-MCP tools once per workspace. The control plane then exposes these integrations to each machine's agent through a single built-in MCP (Machine Control Protocol) server. Agents can discover and call these tools using ocm.search_tools and ocm.call_tool without needing per-integration wiring.

Why Choose OpenClaw Machines? Benefits for the Enterprise

OpenClaw Machines offers compelling advantages:

• Uncompromised Security: With one Firecracker microVM per agent, behind a KVM hardware boundary, security is paramount. Authentication is enforced at the Cloudflare edge and again inside every VM.
• Predictable Cost Efficiency: Instead of paying per agent or per instance, you pay a flat server cost for bare metal. You can run as many hardware-isolated agents as your server can fit, leading to significantly lower per-agent costs at scale. The ability to route to self-hosted LLMs further cuts token expenses.
• True Sovereignty: Your hardware, your data, your keys. The entire platform, from control plane to worker agents, runs on infrastructure you own, giving you complete control and peace of mind regarding data residency and privacy.
• Enterprise Features: Designed with organizations in mind, it includes multi-user accounts and teams, admin-gated host management, encrypted per-machine secrets, and policies for capacity and placement across your fleet.
• Open Source Foundation: Built on Apache-2.0 licensed code, it fosters transparency, community contribution, and allows for deep customization and embedding within your existing systems.

Comparative Advantage: Beyond Traditional Hosting

When evaluating infrastructure options for OpenClaw agents, OpenClaw Machines shines for use cases requiring robust isolation and cost-effective scalability. While local setups and VPS are cheap to start, they lack isolation and don't scale well. Managed services offer ease of use but come with linear, high costs per agent and give up hardware/data control. OpenClaw Machines strikes a balance: it requires a bit more initial setup (provisioning and enrolling a host) but delivers superior economics and hardware-level isolation once you need to run more than a handful of agents. It's about owning your infrastructure and scaling efficiently without compromise.

Getting Started: Prerequisites and Your First Agent

To run OpenClaw Machines, your hosts must be KVM-enabled Linux systems – typically bare metal or a cloud VM with nested virtualization enabled. It explicitly does not support macOS, Windows/WSL, or standard cloud VMs without nested virtualization. A make preflight command helps you check your host's compatibility.

The Getting Started guide is structured in three stages: a local evaluation setup, a production-shaped deployment with Cloudflare and a dedicated host, and finally, exploring the full workflow of creating and using machines, managing their lifecycle, backups, and runtime upgrades. You can even point a coding agent at the documentation to follow the guide.

Practical Takeaways

OpenClaw Machines provides a robust, secure, and cost-effective solution for enterprises looking to scale their AI agent deployments. By leveraging bare-metal servers and Firecracker microVMs, it offers a compelling alternative to managed services and less isolated hosting options. It's ideal for organizations prioritizing data sovereignty, security through hardware isolation, and predictable costs while maintaining an open-source and extensible platform.

FAQ

Q: What kind of isolation does OpenClaw Machines provide for agents?

A: OpenClaw Machines provides hardware-level isolation. Each agent runs in its own Firecracker microVM, which utilizes the KVM hypervisor to ensure a separate guest kernel and a strong hardware boundary, making it safe for untrusted code.

Q: Can I use my own locally served Large Language Models (LLMs) with OpenClaw Machines?

A: Yes, absolutely. The per-host LLM proxy (LiteLLM) is designed to route model traffic not only to various providers but also to models served on your own GPU hardware. This allows for reduced token spend and greater control over your AI inference.

Q: What are the minimum hardware requirements for an OpenClaw Machines host?

A: OpenClaw Machines requires a KVM-enabled Linux host. This can be a bare-metal server or a cloud virtual machine where nested virtualization is explicitly enabled. It will not run on environments like macOS, Windows/WSL, or standard cloud VMs without nested virtualization support.