1. Pick the right host for your workload

Text-only agents (Telegram, WhatsApp, support) run fine on a $6 VPS — Hostinger or Vultr. Agents that call hosted LLM APIs (OpenAI, Anthropic) only need CPU and RAM, not a GPU. Agents that run a local LLM need a GPU host like RunPod or Vast.ai.

2. Install the OpenClaw runtime

On Ubuntu 22.04 or 24.04, install Docker, pull the OpenClaw container, mount a persistent volume for memory, and expose the agent port behind a reverse proxy with TLS. The official installer handles all of this in under five minutes.

3. Wire up your models and tools

OpenClaw supports OpenAI, Anthropic, Google Gemini, local Llama and Qwen via Ollama or vLLM, plus a tool plugin system for HTTP, browser automation, shell, code execution and database access. Every tool runs sandboxed by default.

4. Deploy your first agent

Define an agent in YAML — name, system prompt, allowed tools, memory backend — and OpenClaw provisions a webhook endpoint, a Telegram or Discord adapter, and a memory store. Total time from blank server to a running agent is typically 20-30 minutes.

5. Scale and observe

OpenClaw ships with built-in tracing, token accounting and per-tool latency metrics. Scale horizontally by running multiple agent workers behind a load balancer, or vertically by upgrading the VPS plan. Memory is in Postgres or SQLite by default and can move to a managed database without code changes.