Run as many isolated OpenClaw agents as you need, on hardware you own.
OpenClaw Machines is an open-source platform for running
OpenClaw in secure AI sandboxes on your
own infrastructure. A control plane orchestrates your hosts, and each
agent runs in its own
Firecracker microVM on
them — hardware-isolated, safe for untrusted and agent-generated code. A
Cloudflare data plane is the front door: every machine gets its own
subdomain behind edge auth, reached through a tunnel that terminates inside
the VM — no host port is exposed for user-to-VM traffic. The current control
plane still needs private or firewall-restricted access to each agent's
authenticated control API on 9090. See it running at
openclawmachines.com.
The Apache-2.0 public core ships every piece of that stack:
- a minimal control plane — Go API, Postgres-backed accounts, machines, and hosts; placement, machine lifecycle, host enrollment, backups, and durable workflows;
- the host agent
(
ocm-agent) — boots, supervises, and reaps Firecracker microVMs on your enrolled Linux boxes, managing bridge/TAP networking and rootfs staging; - a per-host LLM proxy (LiteLLM) — one place for model keys and BYO-key support, with per-machine usage tracking across providers (or your own locally served models);
- the OpenClaw runtime — the in-VM pieces: auth proxy, web-chat gateway, live terminal, and the artifact-driven runtime staging/upgrade flow;
- the browser runtime — paired Chromium browser VMs with CDP routing and a watchable live view;
- workspace integrations / native MCP — GitHub, Google Workspace, OpenAPI, GraphQL, and remote-MCP tools connected once per workspace and exposed to machines through the OCM MCP facade;
- and the build pipelines that assemble it all — every component's build command, the GCS artifact bucket layout, host provisioning scripts, and the release lanes.
The ocm CLI lives in the separate
mathaix/ocm-cli Apache-2.0 repository.
Click the screenshot to watch the 43-second demo on YouTube. This is a linked image, not an embedded player.
The demo covers host onboarding, agent spin-up, the running Firecracker VM terminal, workspace MCP integrations, and an agent tool call end to end.
- Security. Real isolation, not containers: one Firecracker microVM per agent, with its own guest kernel behind a KVM hardware boundary — and auth enforced at the edge and again inside every VM.
- Cost. One flat server cost: rent a single bare-metal box and run as many hardware-isolated agents as it fits — see how the options compare. The same architecture cuts token spend too: route agents to open-source models running on your own GPU hardware instead of paying per-token APIs.
- Sovereignty. Your hardware, your data, your keys. Run the control plane and workers on machines you own, and route model traffic through the per-host LLM proxy to any provider — or to models served on your own GPUs.
- Open source. Apache-2.0 public core and companion
ocmCLI, permissively licensed for adoption, embedding, and contribution. - Enterprise. Multi-user accounts and teams, admin-gated host management, encrypted per-machine secrets, and capacity/placement policies across your fleet.
- Ecosystem. Browser VMs for web automation, live terminal and web chat, per-VM routing, workspace-scoped native MCP integrations, backups/snapshots, agent memory, and observability with OpenTelemetry/Opik tracing and per-machine usage tracking.
If you run OpenClaw today, you have a few options:
- Local hardware — run it on your own laptop or desktop.
- A VPS (e.g. Hostinger, DigitalOcean) — rent a virtual server and run it there.
- A managed service (e.g. KiloClaw) — spin up a hosted OpenClaw instance and pay per instance.
OpenClaw Machines is the fourth option: rent one bare-metal server (OVHcloud, Hetzner, …), point OpenClaw Machines at it, and spin up as many hardware-isolated OpenClaw instances as the box will hold. One agent or fifty — the cost stays one flat server.
| Feature | Local hardware | VPS (Hostinger) | Managed (KiloClaw) |  |
|---|---|---|---|---|
| Setup effort | Low | Medium | Lowest | Medium (provision + enroll host) |
| Per-agent isolation | Process-level | Shared-kernel / container | Per instance (managed) | Hardware — Firecracker microVM |
| Run many agents | Limited by your box | Limited by VPS size | Yes — but pay for each | Yes — as many as the server fits |
| Multi-user / teams | No | Manual | Varies | Yes — built-in accounts & teams |
| Cost model | Your own hardware | Pay per VPS | Pay per instance | Pay per server (flat) |
| Cost at scale | Doesn't scale | Rises with size | Highest (linear per agent) | Lowest per agent |
| Hardware control | Full (but limited) | Virtualized, shared | None | Full — dedicated bare metal |
| Data & keys stay yours | Yes | Mostly | No (their infra) | Yes — your hardware |
| Backups / snapshots | Manual | Provider snapshots | Managed | Built-in |
| Ops / maintenance | You | You | None | You (self-hosted control plane) |
In short: the managed route is easiest but priced per agent; local and VPS are cheap to start but don't isolate or scale well. OpenClaw Machines trades a little more setup for the best economics and isolation once you're running more than a couple of agents — one server, many hardware-isolated agents, all yours.
OpenClaw Machines turns your own Linux servers into a pool of secure, on-demand sandboxes. Each sandbox is a real Firecracker microVM (its own kernel, hardware-isolated via KVM) that runs one AI agent. The platform is the control plane that creates those VMs, keeps track of them, routes traffic to them, and tears them down — so you can run many untrusted agents safely on infrastructure you own. Think: a mini-cloud for AI agents, that you self-host.
- Control plane (Go backend) — the brain. Accounts, machines, hosts, and config; the API the UI/CLI call; placement and lifecycle orchestration.
- Hosts + worker agents — your Linux boxes. Enroll a host with an install script; its worker agent boots and stops Firecracker microVMs when told to.
- Machines — one isolated microVM per agent. Inside: the OpenClaw agent, a web chat gateway, and a live terminal.
- Browser VMs — separate microVMs running headful Chromium with a live view, driven by the agent over CDP for browser automation.
- Routing / data plane — every running VM gets its own subdomain and a Cloudflare Tunnel that terminates inside the VM, with auth enforced at the edge and again in-VM.
- Workspace integrations (native MCP) — connect external tools once per
workspace (GitHub, Google Workspace, or any OpenAPI / GraphQL / remote-MCP
endpoint); the control plane exposes them to each machine's agent through a
single built-in MCP server, so the agent discovers and calls them with
ocm.search_tools/ocm.call_toolinstead of per-integration wiring.
flowchart TB
U["you — browser / ocm CLI"] --> EDGE["Cloudflare edge<br/>Access auth · Worker route lookup (KV)"]
EDGE -->|dashboard / API| CP["Control plane (Go)<br/>accounts · machines · hosts<br/>placement · lifecycle · backups"]
CP --- DB[("Postgres")]
CP -->|enroll · heartbeat · boot/stop :9090| H1["Host 1 — your Linux box<br/>ocm-agent · LLM proxy · CDP proxy"]
CP -->|…| HN["Host N"]
EDGE -->|per-VM tunnel, terminates inside the VM| VM1
subgraph H1X["Host 1's microVMs"]
VM1["Machine — Firecracker microVM<br/>OpenClaw agent · web chat · terminal<br/>authproxy + cloudflared inside"]
BVM["Browser VM<br/>headful Chromium · live view"]
VM1 -->|CDP| BVM
end
H1 --- H1X
The full design — data plane, routing, tunnels, lifecycle, config, and the build/release flow — is in docs/architecture.md, and the five-layer stack (React UI → Cloudflare edge → Go control plane → host agents → Firecracker sandboxes) is in docs/tech-stack.md.
OpenClaw Machines runs Firecracker microVMs, which require KVM. You need a KVM-enabled Linux host: bare metal, or a cloud VM with nested virtualization enabled. It does not run on macOS, Windows/WSL, or a standard cloud VM without nested virtualization.
Check your host:
make preflightThe Getting Started guide is three stages, each ending with something working:
Using a coding agent? Point it at
docs/getting-started.mdand ask it to follow the guide from Stage 1.
- Local evaluation — the full stack + a real Firecracker machine on one KVM-capable Linux box. No Cloudflare or public domain is required; use an existing KVM host or the optional GCP provisioning example.
- Cloudflare + a dedicated host — the production-shaped deployment: domain, tunnels, edge auth, and an enrolled cloud or bare-metal host.
- The full workflow — create and use machines (chat, terminal, browser VMs), lifecycle, backups, runtime upgrades.
- Getting Started — the three-stage guide above
- User guide — using a machine day-to-day (model, chat, terminal, browser VM, files, logs, traces, backups)
- Workspace integrations / native MCP — connect GitHub, Google Workspace, OpenAPI, GraphQL, and remote-MCP tools once per workspace
- Architecture — data plane, routing, tunnels, lifecycle, workspace integrations / native MCP
- Tech stack — the five layers, client to sandbox
- Local and BYO-host setup
- Control plane deployment profiles
- Self-hosted control plane prerequisites
- LLM operator runbook
- Public docs inventory
- Contributing · Security policy · Code of conduct
ocmCLI project
- GitHub Discussions — questions, ideas, show & tell
- Issues — bugs and feature requests
- Roadmap — the open-source readiness tracker: what's done, what's next
- Found a vulnerability? See the security policy.
See CONTRIBUTING.md and the code of conduct.