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

Hey — if you've been staring at a VPS picker trying to figure out whether 2GB RAM is "good enough" for OpenClaw, or if you should spring for 8GB, I've been there. And I've made the mistake of under-speccing this twice.

I'm Hanks. I test AI agent setups in real workflows, and the number of times I've seen "JavaScript heap out of memory" on an undersized instance is frankly humiliating at this point. This guide is the spec sheet I wish I'd had on day one — built from community data, real-world deployments, and production case studies from people running OpenClaw at scale.

The question I kept asking: what's the actual minimum to stop fighting the hardware and start getting work done?

Here's everything you need.

Overview — What OpenClaw Actually Runs On

OpenClaw is a Node.js process. That's the Gateway daemon sitting in memory 24/7, handling your 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 process that needs to stay alive while you sleep.

RAM is the primary constraint. CPU matters less, disk matters less. Swap helps, but swap does not save you. If you're allocating RAM for OpenClaw, start from that principle.

Minimum vs. Recommended vs. Production Specs

Anything below 2 GB will crash during onboarding or gateway startup. That's the hard floor. Everything above it is a question of how much pain you want to tolerate.

CPU Requirements

Single-Agent vs. Multi-Agent Load

OpenClaw's Gateway is mostly I/O bound — it waits on the LLM response, then dispatches tool calls. Single-agent setups with cloud models (Claude, GPT, DeepSeek) put almost no CPU pressure on your host. The CPU only really wakes up during tool execution: file operations, subprocess calls, browser automation.

Where CPU starts mattering: concurrent agent sessions, heavy cron jobs firing simultaneously, or browser automation running Playwright. Each headless Chromium session adds real CPU load, and if you're running three of those in parallel, a single vCPU starts showing lag.

Recommended Core Counts

For high-volume production deployments or deployments requiring consistent CPU performance — browser automation at scale especially — a dedicated vCPU VPS is recommended over shared vCPU plans.

RAM Requirements

Base OpenClaw Memory Footprint

The Node.js Gateway process itself is the baseline. The absolute minimum is 2 GB RAM, and the recommended is 4 GB RAM. In practice the Gateway process idles at roughly 300–500 MB. The rest of the RAM headroom absorbs the Control UI, in-memory session state, active tool calls, and the Playwright browser process if you're using browser automation.

Browser automation needs 8 GB. That's not a soft guideline — each Playwright browser instance eats 1–2 GB on its own, and if you're running browser skills regularly, 4 GB is too thin.

Rule of thumb without local LLMs:

• 2 GB — functional but fragile. Expect crashes under any non-trivial load.
• 4 GB — the practical daily minimum for a real agent with external APIs.
• 8 GB — stable for most single-agent production setups including browser automation.

Adding Local LLM (Ollama) — Extra RAM Needed

This is where spec requirements jump sharply. OpenClaw itself barely touches your RAM; the model you load into Ollama is where the gigabytes go.

OpenClaw's system prompt alone is 17,000 tokens. Add sub-agent context, tool definitions, and conversation history, and you need 32K context minimum — 65K+ for production with sub-agents. That context eats RAM through the KV cache, on top of the model weights.

Practical RAM requirements when running Ollama on the same machine as OpenClaw:

Small models — 7–8B — hallucinate tool calls and produce format errors. 14B is marginal. The official Ollama–OpenClaw docs recommend a context window of at least 64K tokens for local models used with OpenClaw. Factor that into your RAM math before you commit to a spec.

If your machine has less than 32 GB unified RAM or VRAM, the cleaner move is running OpenClaw with a cloud LLM (Claude, GPT, DeepSeek) and keeping the local hardware footprint small.

Disk / Storage Requirements

Base Install Size

Disk is the least constrained variable. The OpenClaw install itself is light — Node.js, the package dependencies, and the binary. Plan for about 2–3 GB for the install and initial dependencies.

What you actually need to size for:

Logs, Sessions, and Skill Storage Growth

This is where deployments get surprised at month two. OpenClaw writes session transcripts to JSONL files under ~/.openclaw/sessions/, and those grow continuously. Memory files, cron logs, and skill execution logs all accumulate alongside them.

A typical baseline is Ubuntu 22.04 LTS, 4 GB RAM, 1–2 vCPU, and 40 GB NVMe storage with disciplined log retention. "Disciplined log retention" is doing a lot of work in that sentence. Without a log rotation policy, a long-running agent can generate several gigabytes of JSONL per month depending on message volume.

Set up a cron job to prune old session files, or mount a larger volume from the start. NVMe over HDD — the random read/write pattern of session access is noticeably faster on NVMe.

GPU — Optional but Helpful for Local LLMs

When a GPU Makes a Real Difference

If you're running Ollama on the same box as OpenClaw, a GPU isn't optional — it's the difference between an agent that responds in 2 seconds and one that responds in 40.

An 8GB card delivers 40+ tokens per second with 7–8B models at Q4_K_M — fast enough for real work. For an interactive agent responding to Telegram or WhatsApp messages, that's the minimum acceptable speed. Below 25 tok/s, the conversational feel breaks down and users start resending messages assuming the agent didn't receive them.

NVIDIA GPUs have the best Ollama support. AMD works via ROCm but with some rough edges. Apple Silicon unified memory (M-series Macs) handles local LLMs extremely well — the 32 GB Mac Mini at $1,199 is the most popular single-machine setup among the community for running both OpenClaw and a capable local model.

Running CPU-Only with Quantized Models

No GPU? Not a deal-breaker if you're patient or not using local models.

CPU inference via Ollama is 5–10x slower than GPU. On a modern 4-core VPS or Intel NUC, a 7B quantized model generates 3–8 tok/s. For an async agent that handles email triage or file tasks — where nobody's watching a typing indicator — that's actually fine. For interactive conversation, it's painful.

The honest take: if your hardware can't do local inference at a usable speed, use a cloud LLM. DeepSeek's API is $0.28 per million input tokens. The cost to run a real agent workload on cloud APIs is usually $1–5/month. That's cheaper than the electricity to run a GPU box, and the results are better.

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 RAM)

~$4–8/month — the realistic entry point for a stable daily agent.

Oracle's Always Free tier is the community favorite for zero-cost deployment — OpenClaw doesn't need much: 2 OCPU (4 vCPU) + 4 GB RAM works great, or you can max it out at 4 OCPU + 24 GB — all free. The catch: always upgrade to Pay As You Go to prevent idle reclamation. Without it, some users report accounts terminated without warning.

Good for: single agent, external cloud LLM, Telegram or WhatsApp channel, light skill usage. Not good for: browser automation, local LLMs, multiple concurrent agents.

Standard Tier (4 vCPU / 8 GB RAM)

~$8–20/month — the practical sweet spot for most serious users.

Recommended production sizing starts at 8 GB RAM and 2 vCPU for better response stability under concurrent tasks. At this tier you can comfortably run browser automation skills, multiple messaging channels, and heavier cron schedules without memory pressure.

Hetzner is the most-cited provider in the OpenClaw community for European users. Good network, competitive pricing, and solid support documentation.

Power Tier (8+ vCPU / 16+ GB RAM)

~$30–100/month — for multi-agent setups, heavy automation, or split architecture with a remote Ollama inference server.

At this tier you're either running complex multi-agent workflows, browser automation at scale, or pointing OpenClaw at a self-hosted Ollama instance on the same box. For dedicated server deployments, providers to consider include Hetzner's Robot marketplace (excellent European pricing), OVHCloud (global coverage), and Vultr bare metal (US-focused).

The split architecture approach is also worth considering here: run the OpenClaw Gateway on a cheap 4 GB VPS, and point it at a separate machine — a Mac Mini, a GPU workstation, or a dedicated inference server — for local model inference. The Gateway itself is lightweight; separating it from inference work keeps both layers stable.

At Macaron, the infrastructure layer is handled for you — no Gateway sizing, no disk monitoring, no PM2 config. If you want to test whether your workflows actually run before committing to a VPS setup, try it free at macaron.im and run a real task to see how it fits.

Frequently Asked Questions

Q: Can I run OpenClaw on a Raspberry Pi?

Technically yes. A Raspberry Pi 5 with 8 GB RAM can run the OpenClaw Gateway without problems — it's just a Node.js process. The constraint is local LLM inference: the VideoCore GPU has minimal AI acceleration and CPU-only inference at 7B+ models is extremely slow. If you're using cloud APIs (Claude, GPT, DeepSeek), the Pi works fine as a 24/7 always-on host. Expect response times of 3–8 seconds. Max 1–2 concurrent users.

Q: Is 2 GB RAM really enough to start?

It's enough to install and test. It's not enough for reliable daily use. This configuration is only suitable for testing OpenClaw, learning the CLI, and short sessions — expect instability with long-running agents, heavy prompts, multiple tools, or the Control UI under load. Treat 2 GB as a proof-of-concept floor, not a production spec.

Q: Do I need a GPU to run OpenClaw?

No. OpenClaw itself is a Node.js process — it needs no GPU. A GPU only matters if you're running a local LLM via Ollama on the same machine. Using cloud APIs (Claude, GPT, DeepSeek) requires no GPU at all and works fine on any VPS.

Q: What's the minimum spec to run Ollama alongside OpenClaw on the same box?

8 GB RAM minimum to run both comfortably with a 7–8B quantized model. 16 GB for a 14B model. 32 GB+ for a 32B model. The Ollama VRAM requirements guide has a full breakdown including KV cache calculations at different context lengths.

Q: How much disk does OpenClaw use over time?

The install is ~2–3 GB. After that, growth is driven by session JSONL logs and memory files. A moderately active agent might generate 1–3 GB of log data per month. Set up log rotation from day one — prune ~/.openclaw/sessions/ files older than 30–60 days, depending on how much history you need.

Q: What cloud provider does the community use most?

Hetzner for Europe and general VPS use. Oracle Cloud's Always Free tier for zero-cost deployments. DigitalOcean for its one-click OpenClaw marketplace image, which is the easiest starting point for anyone who doesn't want to configure from scratch.

Specs verified against community benchmarks and provider documentation as of February 2026. VPS pricing reflects published rates and may change. Oracle Always Free tier subject to availability by region.