OpenClaw

Automates the deployment and configuration of OpenClaw on Proxmox LXC containers or any Debian/Ubuntu machine.

What You Get

One command creates a ready-to-use OpenClaw environment:

• Ubuntu 24.04 LTS base with Node.js 22.x and build-essential
• Dedicated claw user with passwordless sudo and systemd lingering
• OpenClaw installed with the gateway running as a systemd user service
• Hardened config template (loopback gateway, log redaction, file permissions 600/700, tool policies)
• Memory plugin (memory-lancedb-hybrid) installed post-hatch for persistent semantic + keyword memory
• Tailscale installed and ready for authentication
• Prompt injection defense and sub-agent delegation rules in agents.md.tpl (applied after hatching)
• Git-tracked ~/.openclaw/ config directory for rollback
• Automated daily backups (3 AM, 7-day retention) and 30-day memory cleanup cron
• System-wide PATH via /etc/profile.d/openclaw.sh (survives dotfile replacement)
• Node compile cache and OPENCLAW_NO_RESPAWN for faster CLI starts on LXC/VM hosts

After install, a post-install wizard handles everything that needs human input: AI providers, model roles, API keys, Telegram bot, Tailscale auth — then seeds BOOTSTRAP.md, swaps to an OpenAI model, launches the TUI to hatch the bot, and restores the production model afterward.

Quick Start

Proxmox Host (creates LXC + installs)

bash -c "$(curl -fsSL pveClaw.ivantsov.tech)"

Dialog-based TUI with Simple (recommended) and Advanced modes. Creates an unprivileged LXC with nesting, injects /dev/net/tun for Tailscale, then runs the install script inside.

Existing Linux Machine (installs directly)

bash -c "$(curl -fsSL setupClaw.ivantsov.tech)"

Standalone mode on any Debian/Ubuntu machine. Templates are fetched from GitHub. Must be run as root.

Local Clone

git clone https://github.com/Exploitacious/OpenClaw.git
cd OpenClaw
bash openclaw.sh          # On Proxmox host — creates LXC + installs
# or
sudo bash openclaw-install.sh  # On any machine — installs directly

After Installation

1. Reboot the container to load PATH and services: reboot
2. SSH in: ssh claw@<container-ip> (password: openclaw)
3. Change the default password: passwd
4. Run the post-install wizard:

bash ~/OpenClaw/openclaw-postinstall.sh

The wizard walks through six steps interactively:

Step	What it does
AI Providers	Menu to add Anthropic, Gemini, OpenAI, OpenCode Zen/Go, Ollama (local or remote via OpenWebUI), DeepSeek, xAI, Mistral, OpenRouter, Together, LiteLLM, or any custom OpenAI-compatible endpoint
Embeddings	OpenAI API key for memory search (text-embedding-3-small) — separate from model providers
Model Roles	Choose "OpenCode Go 4 Mains" (current models), "OpenCode Go Future" (MiMo-V2 when available), or manually assign models per role: primary, fallbacks, sub-agents, heartbeat
Telegram	Bot token from @BotFather + your Telegram user ID for DM access
Tailscale	Authentication + Tailscale Serve on port 18789
Finalize	Hook enablement, gateway restart, openclaw doctor --fix, git commit
Hatch	Lets you choose a hatch model, seeds BOOTSTRAP.md, launches TUI with "Wake up, my friend!", then restores production model and installs memory plugin

All steps detect existing config and skip what's already done. Re-run safely at any time. At the end, the wizard handles the hatching process directly — no openclaw onboard involved. It presents a model selection menu for the hatch (GLM-5 recommended, Claude Sonnet for best quality, GPT-4o as fallback, or keep your current model), copies BOOTSTRAP.md from the shipped OpenClaw template into the workspace, and launches openclaw tui --message "Wake up, my friend!". The bot follows the BOOTSTRAP.md instructions to interactively build its personality: SOUL.md, IDENTITY.md, and USER.md. When done, the bot deletes BOOTSTRAP.md on its own. After you exit the TUI, the script restores your production model and installs the memory plugin.

Note: The memory plugin and all workspace content are intentionally NOT created before hatching. Any content in the workspace directory can trigger OpenClaw's hasUserContent check, which skips BOOTSTRAP.md auto-seeding and blocks the hatching dialogue. Our scripts bypass this entirely by seeding BOOTSTRAP.md manually and launching TUI directly.

Scripted / Non-Interactive Mode

For automation or repeatable deployments, pass everything as flags:

bash ~/OpenClaw/openclaw-postinstall.sh \
  --provider opencode-go --provider-key sk-... \
  --provider ollama --ollama-url https://openwebui.example.com \
  --primary-model opencode-go/kimi-k2.5 \
  --fallback-models "opencode-go/minimax-m2.7, opencode-go/glm-5" \
  --heartbeat-model opencode-go/minimax-m2.5 \
  --openai-key sk-... \
  --telegram-token 123456:ABCdef... \
  --telegram-user-id 5361915599 \
  --tailscale-auth-key tskey-auth-... \
  --non-interactive

Use --provider / --provider-key pairs — repeat for each provider. Ollama supports both local installs and remote proxies (OpenWebUI) via --ollama-url with optional API key. In non-interactive mode, hatching is skipped — run the wizard interactively afterward to hatch, or seed BOOTSTRAP.md manually and launch openclaw tui --message "Wake up, my friend!". Run bash openclaw-postinstall.sh --help for all flags.

File Structure

OpenClaw/
├── openclaw.sh              # Proxmox host script (creates LXC + runs install)
├── openclaw-install.sh      # Install script (standalone or via Proxmox host)
├── openclaw-postinstall.sh  # Post-install wizard (providers, models, Telegram, hatch)
├── openclaw-reset.sh        # Reset agent (wipe memory/sessions/personality, re-hatch)
├── templates/
│   ├── openclaw.json.tpl    # Config template (gateway token auto-generated)
│   ├── soul.md.tpl          # Agent personality scaffold
│   └── agents.md.tpl        # Behavioral rules + prompt injection defense
└── README.md

What Gets Installed Where

Path	Purpose
~/.openclaw/openclaw.json	Main config (mode 600)
~/.openclaw/.env	API keys for providers and embeddings (mode 600)
~/.openclaw/agents/main/agent/auth-profiles.json	Registered provider credentials
~/.openclaw/workspace/SOUL.md	Agent personality (generated by hatching process)
~/.openclaw/workspace/AGENTS.md	Behavioral rules (generated by hatching, customize after)
~/.openclaw/workspace/USER.md	User context for personalization (created post-hatch)
~/.openclaw/workspace/skills/memory-lancedb-hybrid/	Memory plugin (installed post-hatch by wizard)
~/.config/systemd/user/openclaw-gateway.service	Gateway systemd service
/etc/profile.d/openclaw.sh	System-wide PATH for openclaw CLI
~/bin/backup-openclaw.sh	Daily backup script
~/backups/	Backup tarballs (7-day retention)

The daily backup includes openclaw.json, .env, credentials/, workspace/, and all auth-profiles.json files (provider credentials). Retained for 7 days.

Customizing Templates

Edit files in templates/ before running the install script to change what ships with every new container.

openclaw.json.tpl

Config template with placeholder tokens:

• __GATEWAY_TOKEN__ — auto-replaced with random hex during install
• __TELEGRAM_BOT_TOKEN__ — set by the post-install wizard or manually

Default model config uses OpenCode Go ($10/month subscription) with a tiered architecture designed to spread rate limits across independent model pools. The "4 Mains" template ships as the default, using only the 4 models currently available:

Role	Model	Rate Limit	Purpose
Primary	Kimi K2.5	~9,250/mo	Main conversation + reasoning (strongest available)
Fallback 1	MiniMax M2.7	~70,000/mo	High-headroom fallback, keeps lights on
Fallback 2	GLM-5	~5,750/mo	Secondary fallback for provider diversity
Sub-agents	MiniMax M2.7	~70,000/mo	Grunt work — research, tool calls, delegation
Heartbeat	MiniMax M2.5	~100,000/mo	Background pings, cheapest possible

A "Future" template is also available for when MiMo-V2-Omni (multimodal, ~10,900/mo) and MiMo-V2-Pro (1M context, ~6,450/mo) launch on OpenCode Go.

Concurrency capped at 2 main / 3 sub-agents to avoid rate-limit walls on subscription throttling. Context pruning with 6h TTL, compaction flush at 40k tokens. Sub-agents use a separate model pool from the primary to preserve conversational quota. The agent is instructed (via AGENTS.md) to escalate to heavier models for complex tasks when available.

Default security posture (single-operator, personal assistant behind Tailscale):

Setting	Value	Meaning
tools.exec.security	full	All commands allowed (no allowlist)
tools.exec.ask	off	Bot runs commands without prompting
tools.exec.host	gateway	Commands run on the host, not sandboxed
agents.defaults.elevatedDefault	full	Elevated operations (sudo, rm, etc.) run without approval
gateway.bind	loopback	Gateway only reachable via localhost + Tailscale Serve
gateway.auth.allowTailscale	true	Tailscale connections trusted

Note: tools.exec.security: "full" means full trust, not full restrictions. It is the most permissive setting. If the bot claims it needs approval to run commands, it is misreading the config — no changes are needed.

soul.md.tpl

Agent personality scaffold. Define who the agent is, its communication style, and safety guardrails.

agents.md.tpl

Behavioral rules, prompt injection defense, and sub-agent delegation strategy. Ships with detection patterns for common attacks (instruction override, encoded payloads, typoglycemia, social engineering) plus tiered sub-agent guidance: default tier (MiniMax M2.7 for grunt work), escalation tier (MiMo-V2-Pro for complex reasoning / large context), and rate-limit awareness rules.

Useful Commands

openclaw doctor --fix       # Health check and auto-fix
openclaw gateway status     # Gateway service info
openclaw logs --follow      # Real-time logs
openclaw tui                # Terminal UI
openclaw configure          # Interactive config wizard
openclaw skills list        # Available skills
openclaw security audit     # Security posture check

Shell aliases (added by install):

openclaw-update   # Update OpenClaw + restart gateway
openclaw-logs     # Shortcut for openclaw logs --follow
openclaw-status   # Shortcut for openclaw gateway status
openclaw-backup   # Run backup now

Resetting the Agent

To wipe the agent's memory and personality without reinstalling OpenClaw:

bash ~/OpenClaw/openclaw-reset.sh

This wipes memory (vector store + markdown logs), sessions (conversation history), and workspace identity files (SOUL.md, AGENTS.md, IDENTITY.md, etc.). The memory plugin is temporarily stashed outside the workspace so the hatching process sees a clean state. The script then presents a model selection menu (GLM-5, Claude Sonnet, GPT-4o, Kimi K2.5, current, or custom), seeds BOOTSTRAP.md from the OpenClaw template, and launches openclaw tui --message "Wake up, my friend!" to re-hatch directly — no openclaw onboard involved. After you exit the TUI, it restores the original model and memory plugin. All API keys, provider credentials, gateway tokens, and config settings are preserved throughout.

Options:

bash ~/OpenClaw/openclaw-reset.sh --keep-soul     # Preserve personality, wipe only memory/sessions
bash ~/OpenClaw/openclaw-reset.sh --full           # Also reset openclaw.json to template defaults
bash ~/OpenClaw/openclaw-reset.sh --skip-hatch     # Don't launch onboard after reset
bash ~/OpenClaw/openclaw-reset.sh --confirm        # Skip confirmation prompt (for scripting)

A timestamped backup is created automatically before every reset at ~/backups/openclaw-pre-reset-*.tar.gz.

Requirements

Proxmox mode (openclaw.sh):

• Proxmox VE 7.x or 8.x
• Root access on the Proxmox host
• Internet access from the host and container
• Ubuntu 24.04 LTS template (auto-downloaded if missing)

Standalone mode (openclaw-install.sh):

• Debian/Ubuntu-based Linux (tested on Ubuntu 24.04)
• Root access
• Internet access

Post-install wizard (openclaw-postinstall.sh):

• Run as the claw user (not root)
• OpenClaw must be installed and in PATH