Guide
OpenClaw System Requirements
Most sizing guides quote vendor minimums nobody has measured. The RAM figures below are taken from OpenClaw containers running real user workloads on our fleet, so you can size a VPS against what the agent actually consumes rather than a guess.
Short answer: 1 vCPU, 1 GB RAM and 5 GB disk runs a personal OpenClaw agent. 2 GB is the comfortable figure once the browser sidecar and a few skills are in play. OpenClaw 2026.9.8 needs Node.js 24.16+ (within 24.x) or 26.1+ — Node 22, 23 and 25 are not supported. The official installer sets up Node for you. No GPU is required at any tier.
At a Glance
| Minimum | Recommended | |
|---|---|---|
| RAM | 1 GB | 2 GB |
| CPU | 1 vCPU | 2 vCPU |
| Disk | 5 GB | 10–20 GB |
| Node.js | 24.16 | 24.x LTS |
| OS | Linux / macOS / Windows | Linux x86_64 or arm64 |
| GPU | None | None (models are remote) |
| Network | Outbound 443 | Outbound 443 + inbound for the gateway |
RAM — What We Actually Measured
We sampled resident memory across OpenClaw agent containers serving live traffic on three separate hosts. Every container was running the same 2026.7.1 image with a mix of chat channels and skills enabled.
| Percentile | Resident memory | What it means |
|---|---|---|
| Minimum | 230 MB | Idle agent, chat-only |
| Median | 374 MB | Typical personal bot |
| 75th percentile | 571 MB | Several skills active |
| 90th percentile | 1265 MB | Heavy skill or browser use |
| Maximum | 1411 MB | Peak observed |
Sample of 19 production containers, measured 17 August 2026. The important shape here is the gap between the median and the 90th percentile: a typical agent sits under 400 MB, but the tail is more than 3× that. Sizing a box at exactly the median is how people end up with an agent the kernel kills mid-conversation.
Practical reading: 1 GB works for a single chat-only agent. 2 GB is the number to pick if you want headroom for the p90 case without thinking about it again.
Node.js Version — Read This Carefully
OpenClaw 2026.10.4, the current release as of 17 October 2026, declares this engine range, unchanged since 2026.9.8:
"engines": { "node": ">=24.16.0 <25 || >=26.1.0" }That is two separate supported windows:
- Node 22 and older — not supported any more
- Node 23.x — not supported
- Node 24.16.0 up to (not including) 25 — supported
- Node 25.x — not supported
- Node 26.1.0 and newer — supported
The range moves between releases. OpenClaw 2026.7.1 still accepted Node 22.22.3+, 24.15+ and 25.9+, so older guides that say Node 22 is fine were right at the time. An install on an unsupported version fails at the engine check rather than at runtime, which is why the error usually looks like an npm EBADENGINE complaint rather than an OpenClaw error. The official installers (curl -fsSL https://openclaw.ai/install.sh | bash, or iwr -useb https://openclaw.ai/install.ps1 | iex on Windows) set up a supported Node for you. Otherwise install Node 24 LTS with nvm instead of forcing the install.
# Check what you have
node -v
# Install a supported LTS
nvm install 24
nvm use 24CPU
OpenClaw is I/O bound. It spends nearly all of its time waiting on model responses, channel webhooks, and disk. 1 vCPU is genuinely enough for a personal agent. A second core helps when a browser sidecar renders pages or a skill does CPU-heavy parsing (PDF extraction, video, OCR).
Sustained high CPU on an idle agent almost always means a runaway skill or a crash loop, not an undersized box — check logs before upgrading the plan.
Disk
The install itself is modest; the growth comes from state. ~/.openclaw holds config, chat history, embeddings, and workspace files.
- 5 GB — minimum, chat-only, modest history
- 10 GB — comfortable with a browser sidecar and skill data
- 20 GB+ — long chat history, media handling, or a coding harness
A browser sidecar pulls in Chromium (roughly 400 MB) plus per-session profile data, which is the single largest jump most people hit.
Do You Need a GPU? No.
OpenClaw does not run model inference locally. Model calls go out over HTTPS to whichever provider you configure, so the GPU question simply does not apply. If you want local inference, run Ollama or LM Studio on a separate machine and point OpenClaw at it — the GPU belongs on the inference host, not the agent host.
Operating System Support
- Linux — first-class, x86_64 and arm64. The official container image is Debian 12 based.
- macOS — Intel and Apple Silicon
- Windows — runs natively from the PowerShell installer, and the Windows Hub app can set up a private gateway for you in its own WSL distro; see the Windows install guide. WSL2 is the most Linux-compatible option
- Docker — works anywhere Docker does; see docker compose
- Raspberry Pi — works on arm64 with 2 GB+; expect slower skill installs
Network
- Outbound 443 to your model provider — required
- Inbound to the gateway port (18789 by default) if you access the web gateway remotely
- Inbound 80/443 if you terminate TLS for channel webhooks
Sizing Cheat Sheet
- Personal bot, chat-only: 1 vCPU / 1 GB / 5 GB — any $4–5/mo VPS
- Personal bot + browser sidecar: 1 vCPU / 2 GB / 10 GB
- Team bot, 5–10 active users: 2 vCPU / 4 GB / 20 GB
- Coding harness / heavy skills: 2–4 vCPU / 4–8 GB / 40 GB
Skip the Sizing Question
OpenClaw Launch runs OpenClaw on right-sized containers with the supported Node version, browser sidecar, and channel wiring already in place. Compare self-hosting costs in the cheapest OpenClaw hosting guide or see the OpenClaw VPS guide if you would rather run it yourself.