OpenClaw on a Lean VPS

Complete Optimization Guide for 1 vCPU · 2 GB RAM

🗓 April 20, 2026  ·  🐐 Babu AI  ·  Research Report

Everything you need to run OpenClaw efficiently on a tiny, constrained VPS — without it falling over.

OpenClaw Lean VPS Optimization
01 / 15

The Reality of 2 GB ★

760
MB — Gateway at rest (measured)
400
MB — Each subagent process (measured)
200
MB — OS + buffer cache overhead

The Math Doesn't Lie

Gateway (760 MB) + 1 Subagent (400 MB) + OS (200 MB) = 1.36 GB used

Any burst = LLM token processing spikes to +400 MB → OOM territory

OpenClaw Lean VPS Optimization
02 / 15

TL;DR — The 8 Golden Rules

  1. One channel only — Telegram webhook is ideal
  2. Heap cap: 1536 MB — set --max-old-space-size=1536
  3. Sandbox OFF — it doubles memory footprint
  4. Max 1 concurrent session — no parallel conversations
  5. Max 1 concurrent cron — cron.maxConcurrentRuns=1
  6. Heartbeat: 60 min — not 30, not 15
  7. Skip deep-research — too heavy for 2 GB
  8. Watch MemAvailable — alert below 300 MB
OpenClaw Lean VPS Optimization
03 / 15

1. Channels — Pick One and Stick With It

webchat 10–30 MB · BEST
Telegram (webhook) 20–40 MB · RECOMMENDED
Email IMAP IDLE 15–40 MB · EXCELLENT
Slack (webhook) 40–80 MB · FINE
Discord 50–120 MB · VIABLE
WhatsApp (official API) 20–50 MB · WEBHOOK ONLY
WhatsApp (Baileys / unofficial) 100–250+ MB · AVOID
Signal 150–250+ MB · AVOID
OpenClaw Lean VPS Optimization
04 / 15

Channel Deep Dive

✅ webchat — Best Choice
  • Zero external API calls
  • No persistent WebSocket to manage
  • Scales predictably
  • No rate limits
  • Only downside: requires browser
✅ Telegram — Best External Channel
  • Set webhook mode = no polling
  • Very low RAM footprint
  • Well-documented, stable bot API
  • Requires public HTTPS endpoint
  • Free, unlimited messages
⚠️ Discord — Viable But Heavier
  • Persistent WebSocket gateway connection
  • Moderate memory overhead
  • Best for community bots
  • Avoid on high-member guilds
❌ Signal / WhatsApp Unofficial
  • No official bot API for Signal
  • Baileys = 150–250 MB RAM alone
  • Experimental, unstable
  • Not viable for production
OpenClaw Lean VPS Optimization
05 / 15

2. Skills — Less Is Dramatically More

✅ Fine to Install
  • weather — stateless, tiny
  • taskflow — lightweight coordination
  • healthcheck — on-demand audits
  • summarize — small, focused
❌ Skip These on 2 GB
  • deep-research — 5–8 parallel subagents
  • skill-creator — heavy tooling builds
  • Any browser/Playwright skill — 100–300 MB per browser
  • blogwatcher — persistent subprocess
  • superpowers-* — use one at a time max

Rule: Audit Your Skills

Every skill adds 5–50 MB RAM. Run openclaw skills list and delete anything you haven't used in a week.

OpenClaw Lean VPS Optimization
06 / 15

3. Config — The Critical Knobs

agents.defaults.maxConcurrent 1
agents.defaults.subagents.maxConcurrent 1
agents.defaults.subagents.maxChildrenPerAgent 2
cron.maxConcurrentRuns 1
session.maintenance.pruneAfter "7d"
session.maintenance.maxEntries 100
logging.level "warn"
agents.defaults.sandbox.mode "off"
tools.profile "minimal"
OpenClaw Lean VPS Optimization
07 / 15

3b. Heartbeat & Cron — Slow It Down

Heartbeat Settings
heartbeat.every"60m" (not 30m)
heartbeat.isolatedSessiontrue
heartbeat.lightContexttrue
Cron Settings
cron.maxConcurrentRuns1 (never more)
cron.sessionRetention"1h"
cron.runLog.maxBytes500000
cron.runLog.keepLines500

Why "60 minutes"?

Every heartbeat fires a full agent turn — model inference, context loading, tool calls. On 2 GB, that contention with user conversations is not worth it. 60 minutes is plenty for periodic inbox checks.

OpenClaw Lean VPS Optimization
08 / 15

4. Memory — Cap the Heap First

--max-old-space-size

The single most impactful change on a constrained VPS

❌ Default (often 2048 MB)
On 2 GB total RAM, this leaves virtually nothing for the OS.

Every GC pause and token burst pushes the system into swap or OOM.
✅ Recommended (1536 MB)
Gateway: --max-old-space-size=1536
Subagents: --max-old-space-size=768

Leaves ~500 MB for OS, buffer cache, kernel — critical headroom.
NODE_OPTIONS="--max-old-space-size=1536"
OpenClaw Lean VPS Optimization
09 / 15

4b. System-Level Memory Tuning

Swappiness

sysctl vm.swappiness=10

Default 60 pushes Node heap to swap — catastrophic on VPS disk.

Add permanently:

echo "vm.swappiness=10" >> /etc/sysctl.conf

Swap File

fallocate -l 2G /swapfile && chmod 600 /swapfile && mkswap /swapfile && swapon /swapfile

Emergency headroom only — do NOT treat as usable memory. VPS swap is slow; processes that hit it stall.

Systemd Memory Guard
sudo systemctl edit openclaw-gateway
[Service]
MemoryMax=2G
MemorySwapMax=1G

This makes the kernel prefer killing OpenClaw via cgroup OOM before the system-wide OOM killer fires unpredictably.

OpenClaw Lean VPS Optimization
10 / 15

5. Monitoring — Watch These Numbers

Key Commands

ps aux | grep openclaw # RSS per process
free -m # MemAvailable
dmesg | grep -i oom # OOM events
openclaw gateway status # built-in status
systemctl show openclaw-gateway --no-pager

Alert Thresholds

RAM used > 75%Start investigating
RAM used > 90%Immediate action
MemAvailable < 300 MBCRITICAL
CPU > 80% sustainedCheck cron jobs
Any OOM eventReview immediately

The Most Important Number

MemAvailable — when it drops below 300 MB, you are in the danger zone. Alert immediately.

OpenClaw Lean VPS Optimization
11 / 15

6. Anti-Patterns — The Kill List

Multiple channels simultaneously Run 1 only
Sandbox mode (Docker) Doubles RAM
deep-research skill OOM trigger
cron.maxConcurrentRuns > 1 Parallel OOM spike
WhatsApp Baileys / unofficial libs 150–250+ MB
Browser automation / Playwright 100–300 MB/browser
Large channel history limits Context bloat
Docker install on <2 GB VPS pnpm OOM
Signal bot No official API
OpenClaw Lean VPS Optimization
12 / 15

7. Recovery — When Things Go Wrong

  1. 1 Restart the gateway
    openclaw gateway stop && openclaw gateway start
  2. 2 Confirm OOM:
    dmesg | grep -i oom
    Exit 137 = OOM confirmed
  3. 3 Clear cron logs
    find ~/.openclaw/cron/runs/ -type f -delete
  4. 4 Prune sessions — reduce maxEntries and pruneAfter, restart
  5. 5 Disable all but one channel — edit openclaw.json, restart
  6. 6 Ensure sandbox is OFF — agents.defaults.sandbox.mode: "off"
  7. 7 Clear media cache:
    rm -rf ~/.openclaw/media/
  8. 8 Add swap if missing — safety net for future spikes
OpenClaw Lean VPS Optimization
13 / 15

The Survival Formula ★

1 Channel
+ Sequential Sessions
+ No Sandbox
+ 1536 MB Heap Cap
+ Swap
+ 60-min Heartbeat
+ No Heavy Skills

= Survivable on 2 GB

The tighter you run, the more headroom for the work that actually matters.
Every feature addition requires a compensating adjustment elsewhere.

OpenClaw Lean VPS Optimization
14 / 15

Final Checklist — Confirm Each Item

  1. ☐ Exactly 1 channel (Telegram webhook recommended)
  2. ☐ agents.defaults.maxConcurrent: 1
  3. ☐ agents.defaults.subagents.maxConcurrent: 1
  4. ☐ cron.maxConcurrentRuns: 1
  5. ☐ heartbeat.every: "60m" + lightContext: true
  6. ☐ agents.defaults.sandbox.mode: "off"
  7. ☐ session.maintenance.pruneAfter: "7d"
  8. ☐ session.maintenance.maxEntries: 100
  9. ☐ logging.level: "warn"
  10. ☐ NODE_OPTIONS="--max-old-space-size=1536"
  11. ☐ 1–2 GB swap file configured
  12. ☐ vm.swappiness=10
  13. ☐ Only essential skills installed
  14. ☐ deep-research, skill-creator, browser skills NOT installed
  15. ☐ Alerting: MemAvailable < 300 MB
OpenClaw Lean VPS Optimization · Research by Babu AI for Thota · April 20, 2026
15 / 15