Run OpenClaw on a Dedalus Machine

A cookbook example: spin up a Dedalus Machine (Ubuntu microVM), install OpenClaw, and chat with it from your terminal.

The script is provider-agnostic — point it at any LLM that OpenClaw supports (Anthropic, OpenAI, etc.) by setting LLM_MODEL and LLM_API_KEY.

Prereqs

• Node.js 20+
• A Dedalus API key — sign up at https://www.dedaluslabs.ai/
• An LLM provider API key (Anthropic, OpenAI, …)

Quickstart

cp .env.example .env       # fill in DEDALUS_API_KEY, LLM_MODEL, LLM_API_KEY
npm install
npx tsx openclaw.ts        # creates a machine, installs OpenClaw, sends one chat

Expected end-to-end runtime: ~3 minutes (machine provision ~30s, npm install -g openclaw ~2min, one chat round ~5s).

The script ends by printing the assistant's reply and the machine ID. To send another message to the same machine:

npx tsx chat.ts <machine-id> "what's 2 + 2?"

Choosing a model

LLM_MODEL is sent to the gateway as the x-openclaw-model header, so any model OpenClaw routes to works. Examples:

Provider	LLM_MODEL	LLM_API_KEY
Anthropic	anthropic/claude-sonnet-4-6	sk-ant-...
OpenAI	openai/gpt-4o	sk-...

Any provider OpenClaw lists at https://docs.openclaw.ai works the same way.

Machine sizing

Resource	Default	Why
vCPU	2	One core for the gateway, one for the agent turn
Memory	4096 MiB	Gateway idles ~300 MiB; agent turns spike higher
Storage	10 GiB	Persistent volume mounted at /home/machine

These are the values this example runs at. Smaller machines may work, but npm install -g openclaw on a cold machine is memory-hungry; if you go below 4 GiB, watch for OOM during install.

What openclaw.ts does

1. Create the machine with client.machines.create({ vcpu, memory_mib, storage_gib }) and wait for phase === "running".
2. Install Node.js 22 + OpenClaw under /home/machine/.npm-global (the persistent volume).
3. Configure the gateway — gateway.mode=local, enable /v1/chat/completions, bind to loopback, no auth, mirror LLM_API_KEY into both ANTHROPIC_API_KEY and OPENAI_API_KEY env vars.
4. Start the gateway detached on port 18789.
5. Send one chat message — POST to /v1/chat/completions over the execution API and print the reply.

Notes on security

The example binds the gateway to loopback only (gateway.bind=loopback) and disables auth (gateway.auth.mode=none). All chat traffic flows through the Dedalus execution API, which already authenticates via your DEDALUS_API_KEY.

To expose the gateway to the public internet via a Dedalus preview tunnel, add token auth first — see https://docs.openclaw.ai/web for gateway.auth.mode=token and controlUi.allowedOrigins.

Files

.env.example      DEDALUS_API_KEY, LLM_MODEL, LLM_API_KEY
openclaw.ts       end-to-end: create machine + install + configure + one chat round
chat.ts           send one message to an existing machine
package.json