OpenClaw Hardware Requirements: Minimum and Recommended Specs for Self-Hosting (2026)

Hey — if you've been staring at a VPS picker wondering whether 2 GB of RAM is "good enough" for OpenClaw or whether you should spring for 8 GB, I've been there. I've under-specced this twice.

I'm Hanks. I test AI agent setups in real workflows, and the number of times I've watched "JavaScript heap out of memory" scroll past on an undersized box is frankly humiliating at this point. (If you're seeing that error right now, it's almost always RAM — the troubleshooting guide covers the OOM case.) This is the spec sheet I wish I'd had on day one. The question I kept asking: what's the actual minimum to stop fighting the hardware and start getting work done — and which configs are a trap?

Here's the whole thing as a judgment call, not a spec dump. Last verified June 2026.

The Judgment Table: Pick Your Scenario

Find the row that matches what you're actually going to run. The first column is the trap — specs that technically boot but will waste your time. The single rule that governs everything below: RAM is the constraint, and when you're unsure, size up one tier.

If your row's "minimum" makes you wince, that's the signal to either move up a tier or push inference to the cloud and keep the host small. The rest of this page is the reasoning behind each cell — and where to go once you've picked a spec (install guide for a clean setup, VPS deployment guide if you're renting a box).

Overview — What OpenClaw Actually Runs On

OpenClaw is a Node.js process. That's the Gateway daemon sitting in memory 24/7, handling message routing, session persistence, cron jobs, and tool calls. It's not a chat window you open and close — it's a long-running server that stays alive while you sleep.

That single fact decides everything: RAM is the primary constraint, CPU matters less, disk matters less. Swap helps, but swap does not save you. The heavy AI computation doesn't happen here — OpenClaw is model-agnostic and offloads inference to whatever provider you bring (Anthropic, OpenAI, DeepSeek, or a local model). So unless you're running that model on the same machine, your local footprint stays small.

OS & Runtime Requirements (The Other Half of "Not Suitable")

Before you size RAM, make sure the box can run OpenClaw at all. This is where the most common "won't even start" mistakes live — and most of them are software, not hardware.

Three facts worth pinning down, because they've each cost me an evening:

• Native Windows is not supported — full stop. Install WSL2 with Ubuntu and follow the Linux path inside it. The install guide walks the WSL2 setup; don't fight the native path.
• Node 22 is mandatory. OpenClaw uses modern JavaScript (native fetch, ES module and V8 improvements) that simply isn't in older runtimes, and recent 22.x releases also ship security fixes the Gateway relies on. Pin a current 22.x; the installer detects and warns if you're behind.
• Linux needs glibc 2.31+ — Ubuntu 22.04/24.04, Debian 12, Fedora 39+, and Arch all qualify. Apple Silicon (M1–M4) and Intel x86-64 are both fully supported on macOS 11+.

OpenClaw also ships updates fast (I've watched the version jump several point releases in a single week), and dependency requirements drift with them. Run openclaw doctor after every upgrade to catch configuration drift, and re-check this section after major releases.

RAM Requirements — The One That Actually Bites

The Gateway process idles at roughly 200–500 MB. The rest of your RAM headroom absorbs the Control UI, in-memory session state, active tool calls, and — the big one — the browser process if you're using automation.

Anything at or below 2 GB will crash during onboarding or gateway startup. That's the hard floor, not a pessimistic estimate — I learned it on a 2 GB box where three concurrent scraping agents got OOM-killed in a round-robin loop until I doubled the RAM. Each headless browser session Playwright spins up grabs anywhere from a few hundred MB to over a gigabyte depending on the page, so browser skills are the fastest way to blow past your budget. If you keep hitting memory ceilings, the troubleshooting guide has the OOM and heap-size fixes.

Adding a Local LLM (Ollama) — Where RAM Jumps

OpenClaw barely touches your RAM; the model you load into Ollama is where the gigabytes go. Its system prompt plus sub-agent context, tool definitions, and history push you toward a 32K-context minimum (64K+ for production), and that context eats RAM through the KV cache on top of the model weights.

Small models (7–8B) hallucinate tool calls and produce format errors; 14B is marginal. The official Ollama–OpenClaw docs recommend at least a 64K context window for local models. If your machine has less than 32 GB of unified RAM or VRAM, the cleaner move is a cloud LLM with a small host — my full walkthrough of that trade-off is in the OpenClaw + DeepSeek setup guide.

CPU Requirements

OpenClaw's Gateway is mostly I/O bound — it waits on the LLM response, then dispatches tool calls. Single-agent setups with cloud models put almost no CPU pressure on the host; the CPU only wakes up during tool execution, subprocess calls, and browser automation.

The "not suitable" case here is subtle: a shared 1 vCPU plan running browser automation at scale. For consistent performance under concurrent tasks, a dedicated vCPU VPS beats a shared one every time.

Disk / Storage Requirements

Disk is the least constrained variable — until it isn't. The install itself is small (a few hundred MB of Node modules); budget 2–3 GB once you add Docker images and initial dependencies.

Two "not suitable" traps: a microSD card (on a Raspberry Pi) and an HDD. OpenClaw writes session transcripts to JSONL files under ~/.openclaw/sessions/ continuously, and that random read/write pattern is painfully slow on microSD or spinning disk — use NVMe (or at least a USB SSD on a Pi). And without log rotation, a busy agent can generate several gigabytes of JSONL a month, so set a cron job to prune old sessions from day one or mount a larger volume up front.

GPU — Optional for Cloud, Decisive for Local

If you're using cloud APIs, you need no GPU at all — skip this section. A GPU only matters when you run Ollama on the same box, and there it's the difference between a 2-second reply and a 40-second one.

The clear "not suitable" line: CPU-only inference for anything past ~7B. A 7B quantized model on a modern 4-core CPU manages 3–8 tokens/sec — fine for async email triage where nobody's watching a typing indicator, painful for live chat. A 13B model on CPU is worse, and anything larger is impractical for interactive use. NVIDIA has the best Ollama support; AMD works via ROCm with rough edges; Apple Silicon's unified memory handles local models well. Honestly, if your hardware can't do local inference at a usable speed, use a cloud model — DeepSeek's API runs a real agent workload for roughly $1–5/month, cheaper than the electricity for a GPU box.

Cloud VPS Sizing Guide

Cloud is the right answer for most self-hosters who don't want to manage local hardware. The Basic DigitalOcean plan with at least 4 GB of RAM is recommended for running OpenClaw effectively. That benchmark holds across providers.

Ubuntu 22.04 LTS is the recommended OS for OpenClaw self-hosting — best Docker compatibility, longest support lifecycle, and most community documentation. Debian 12 is a solid alternative.

Budget Tier (2 vCPU / 4 GB) — ~$4–8/mo

Good for a single agent on a cloud LLM with one channel. Not for browser automation, local LLMs, or multiple concurrent agents. Oracle's Always Free tier is the community zero-cost favorite — just switch to Pay As You Go to avoid idle reclamation.

Standard Tier (4 vCPU / 8 GB) — ~$8–20/mo

The practical sweet spot: comfortable browser automation, multiple channels, and heavier cron schedules without memory pressure.

Power Tier (8+ vCPU / 16+ GB) — ~$30–100/mo

For multi-agent workflows, automation at scale, or a split architecture — run the lightweight Gateway on a cheap 4 GB VPS and point it at a separate Mac Mini, GPU workstation, or inference server. Separating the Gateway from inference keeps both layers stable.

At Macaron, the infrastructure layer is handled for you — no Gateway sizing, no disk monitoring, no PM2 config. If you'd rather test whether your workflows run before committing to a VPS, try it free at macaron.im and throw a real task at it.

Frequently Asked Questions

Q: Can I run OpenClaw on a Raspberry Pi? Yes, in cloud mode. A Pi 5 (8 GB) runs the Gateway fine as a 24/7 host when inference is handled by a cloud API — expect 3–8 second response times and 1–2 concurrent users. Use an NVMe/USB SSD, not the microSD slot, for the persistent session writes. Local LLM inference on the Pi is the part that doesn't work: minimal AI acceleration and very slow CPU inference at 7B+.

Q: Is 2 GB RAM really enough to start? Enough to install and test, not enough for reliable daily use. Treat 2 GB as a proof-of-concept floor — expect instability with long-running agents, heavy prompts, multiple tools, or the Control UI under load. 4 GB is the daily minimum.

Q: Do I need a GPU? No — OpenClaw is a Node.js process and needs no GPU. A GPU only matters if you run a local LLM via Ollama on the same machine. Cloud APIs require none.

Q: Why won't OpenClaw start on my Windows machine? Native Windows isn't supported. Install WSL2 with Ubuntu and run the Linux path inside it; the install guide has the steps. If it installed but won't launch, check your Node version (22.12+) and the troubleshooting guide.

Q: What's the minimum to run Ollama alongside OpenClaw on one box? 8 GB RAM with a 7–8B quantized model, 16 GB for 14B, 32 GB+ for 32B — and a GPU if you want interactive speed. Full local-model walkthrough is in the DeepSeek/Ollama guide.

Q: How much disk does OpenClaw use over time? Install is ~2–3 GB; after that, growth comes from session JSONL logs and memory files — a moderately active agent might add 1–3 GB/month. Set up log rotation on day one and prune ~/.openclaw/sessions/ older than 30–60 days.

Q: What cloud provider does the community use most? Hetzner for Europe and general use, Oracle Cloud's Always Free tier for zero-cost, and DigitalOcean for its one-click OpenClaw marketplace image if you don't want to configure from scratch.

Specs verified against community benchmarks and provider documentation as of June 2026. Node, OS, and Docker requirements shift with frequent OpenClaw releases — re-check the install guide for the current pinned versions. VPS pricing reflects published rates and may change; Oracle Always Free availability varies by region.