---

📋 Table of Contents

• 🏁 Start Here — What is OpenClaw?
• ⏳ Timeline & Milestones
• 🏠 Official Resources
• ⚡ Getting Started & Installation
• 🧠 Core Architecture
• 🤖 LLM Model Selection Guide
• 💰 Cost Estimation Guide
• 🔧 Skills & Plugins
• 🤖 MCP Integrations
• 🖥️ UIs & Companion Apps
• 🌐 Platform Channels & Messaging
• ☁️ Deployment & Infrastructure
• 🏢 Enterprise Deployment
• 🧠 Memory & Knowledge Systems
• 🔒 Security
• 🆚 OpenClaw vs Other Agent Frameworks
• 🌍 Ecosystem & Agent Platforms
• 🛠️ Developer Tooling
• 🏗️ Alternative Implementations
• 📖 Tutorials, Blogs & Articles
• 🎥 Videos & Talks
• 📚 Research & Academic Resources
• 💼 Use Cases & Showcases
• ❓ FAQ & Troubleshooting
• 🤝 Community & Support
• 🗺️ Related Awesome Lists
• 🤝 Contributing

---

🏁 Start Here — What is OpenClaw?

OpenClaw (formerly Clawdbot, briefly Moltbot) is an open-source, local-first autonomous AI agent framework. You run a persistent process called the Gateway on hardware you control — a Mac mini, a VPS, a Raspberry Pi — and it connects to messaging apps you already use. When you send a message, OpenClaw runs an agent turn powered by an LLM, invokes tools or skills, and responds — often taking real actions on your behalf.

Key design decisions that explain the hype:

• Messaging app as UI — No new app to install. WhatsApp, Telegram, Discord, Slack, iMessage — whatever you already live in becomes your agent interface.
• Local-first, self-hosted — Your data stays on your machine. No third-party server sees your context.
• Skills/Plugins architecture — Capabilities are tiny markdown-declared plugins. Anyone can write one in under an hour.
• Persistent memory — Flat-file Markdown memory (MEMORY.md, SOUL.md) persists across restarts without a vector database.
• Model-agnostic — Works with Claude, GPT-5, Gemini, local Ollama models, and OpenRouter.

---

⏳ Timeline & Milestones

Understanding OpenClaw's history explains its current architecture and community dynamics.

Nov 2025  ──  Peter Steinberger (@steipete) releases "Clawdbot"
              TypeScript, Claude-only, WhatsApp + Telegram support
              Day 1: 5K stars. Week 1: 30K stars.

Dec 2025  ──  Anthropic trademark dispute → renamed "Moltbot"
              72 hours later renamed again to "OpenClaw"
              Community skill ecosystem begins forming spontaneously
              ClawHub launched as community skill registry

Jan 2026  ──  OpenClaw goes supernova: 68K → 120K stars in 2 weeks
              CVE-2026-25253 disclosed — WebSocket RCE vulnerability
              341 malicious skills discovered on ClawHub
              Alternative implementations start appearing:
              ZeroClaw (Rust), NanoClaw (Python), PicoClaw (Go)

Feb 2026  ──  Steinberger joins OpenAI
              Project transferred to OpenClaw Foundation (independent)
              ClawHub adds VirusTotal scanning for all submitted skills
              NemoClaw (NVIDIA, sandboxed) released
              Moltis (Rust, enterprise-grade) hits 1.0

Mar 2026  ──  247K stars, 47.7K forks
              OpenClaw Foundation announces governance model
              13,729 skills on ClawHub
              v1.9.x stable release series

---

🏠 Official Resources

Monorepo package map:

---

⚡ Getting Started & Installation

Quick Install

# Requires Node.js 20+
npm install -g openclaw

# Interactive onboarding wizard
openclaw onboard

# Or specify your provider directly
openclaw onboard --auth-choice anthropic-api-key
openclaw onboard --auth-choice openai-api-key
openclaw onboard --auth-choice openrouter        # Recommended for cost optimization

Docker (Recommended for Production)

docker run -d \
  -p 127.0.0.1:18789:18789 \
  -v ~/openclaw-data:/data \
  -e ANTHROPIC_API_KEY=your_key \
  --restart unless-stopped \
  openclaw/openclaw:stable

⚠️ Critical: Bind to 127.0.0.1, not 0.0.0.0. See Security for why this matters.

Release Channels

Install Guides

• Official Quickstart — 5-minute path to a working agent
• Docker Setup Guide — Recommended for all production deployments
• Nix/NixOS Setup — Reproducible, declarative via nix-openclaw
• DigitalOcean 1-Click — Fastest cloud deploy
• VPS Hosting & Deployment Guide — Nginx, SSL, firewall
• Raspberry Pi Setup — Runs on Pi 4 4GB+
• Mac Mini as Home Server — Highly recommended homelab
• Ansible Playbook — Automated secure installation with Tailscale VPN, UFW firewall and Docker isolation

---

🧠 Core Architecture

Five concepts unlock the entire ecosystem:

1. The Gateway

A WebSocket server (port 18789) that is the single orchestration point. It normalizes heterogeneous messaging protocols — WhatsApp via Baileys, Telegram via grammY, Slack Events API, Discord gateway — into a canonical internal message format. No intelligence lives here. It is a pure traffic controller.

2. Skills vs Plugins vs MCP

3. Memory Architecture

OpenClaw uses filesystem-as-truth memory — no vector database required by default.

4. Context Engineering

When approaching token limits, OpenClaw triggers session compaction — summarizing oldest turns while preserving tool call-result pairs. Sessions reset at configurable boundaries (default: 4 AM UTC). Result: effectively unlimited conversation length.

5. Channel Isolation

Each connected messaging platform is a channel with its own context namespace. Skills can be scoped per channel using channels: in the skill manifest. Run 10+ channels simultaneously with per-channel permission sets.

---

🤖 LLM Model Selection Guide

This is one of the most common questions new users ask, and no other resource covers it well. Here is the community consensus as of March 2026.

Model Comparison at a Glance

Recommendation by Use Case

Personal productivity agent (email, calendar, tasks): → Claude Sonnet 4 as primary. Claude Haiku 4 for simple reminders and quick lookups. Route with ClawRouter.

Software development assistant: → GPT-5 or Claude Sonnet 4. Both excel at code. GPT-5 slightly stronger on large refactors; Sonnet 4 more reliable on multi-step tool chains.

Research and document processing: → Gemini 2.5 Pro for its 1M token context window. Process entire codebases or large PDFs without chunking.

Privacy-first / air-gapped deployment: → Llama 4 Scout (16GB VRAM) or Qwen3-14B (8GB VRAM) via Ollama. No data leaves your machine.

Cost-optimized high-volume: → Gemini 2.5 Flash or Claude Haiku 4. Use ClawRouter to auto-route cheap tasks to these and expensive tasks to Sonnet/GPT-5.

Configuring Multiple Models

# Configure a primary and fallback model
openclaw config set model.primary "Claude Sonnet 4.6"
openclaw config set model.fallback "gemini-2.5-flash"
openclaw config set model.coding "gpt-5"

# Or use OpenRouter for automatic routing
openclaw config set provider "openrouter"
openclaw config set model.primary "openrouter/auto"

Local Model Setup with Ollama

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull a recommended model
ollama pull llama4:scout        # 16GB VRAM
ollama pull qwen3:14b           # 8GB VRAM
ollama pull mistral-nemo:12b    # 6GB VRAM (minimum viable)

# Point OpenClaw at local Ollama
openclaw config set provider "ollama"
openclaw config set model.primary "llama4:scout"

---

💰 Cost Estimation Guide

One of the top reasons people hesitate to try OpenClaw. Here are real numbers.

API Cost by Provider (March 2026 Pricing)

"Typical turn cost" = one user message + agent response, including tool calls. Complex multi-step tasks cost more.

Monthly Cost Estimates by Usage Profile

Actual costs vary dramatically based on model choice, skill complexity, and conversation length. The cost-tracker skill gives you real per-conversation data.

Cost Optimization Strategies

1. Use ClawRouter for intelligent routing (40–60% savings)

openclaw skill install BlockRunAI/clawrouter
# Routes simple tasks (reminders, lookups) to Haiku/Flash
# Routes complex tasks (code, research) to Sonnet/GPT-5
# Users report 40-60% cost reduction

2. Context window hygiene

# Enable aggressive session compaction
openclaw config set context.compaction_threshold 0.7
openclaw config set context.reset_schedule "0 4 * * *"  # 4am UTC daily

3. Scope skills narrowly Skills with broad triggers (use this when the user asks anything about...) consume more tokens evaluating whether to activate. Narrow trigger conditions = fewer wasted tokens.

4. Use local models for bulk tasks Run Ollama locally for high-volume, lower-complexity tasks (email sorting, routine summaries). Zero marginal API cost.

Total Cost of Ownership: Self-Hosted vs Managed

---

🔧 Skills & Plugins

Official Skill Registry (ClawHub)

clawhub.ai — The official marketplace. 13,729 community-built skills as of March 2026.

openclaw skill install steipete/slack
openclaw skill install community/playwright-mcp
openclaw skill list
openclaw skill remove steipete/slack

⚠️ ClawHub is open-submission. Not all skills are vetted. Always review source before installing. See the Security section.

Curated "Must-Have" Skills

🛡️ Install This First:

Productivity Core:

Development:

Research & Data:

Skills by Category

For the full categorized index of 5,400+ vetted skills: ➡️ VoltAgent/awesome-openclaw-skills — 1M+ monthly views, the #1 community skills resource.

Building Your Own Skills

A skill is a SKILL.md file. Minimum viable skill:

# my-skill

Use this skill when the user asks to [TRIGGER CONDITION].

## Setup
- Requires: API_KEY environment variable

## Tools

### do_thing
Does the thing.
Parameters:
- query (string): What to do

## Instructions
When triggered, call do_thing with the user's request.
Return the result in a friendly format.

Skill authoring resources:

• Official Skills Documentation — SKILL.md loading mechanics, scope priority, token cost formula
• SKILL.md Format Specification — frontmatter fields, env var declarations, binary dependencies, install spec
• ClawHub CLI Publishing Tool — clawhub publish <path>, clawhub inspect <slug>, versioning, soft-delete/restore
• Skills: Install & Custom Development in Practice — Real-world SKILL.md examples (ClickUp, Outlook, GitHub) with minimal permission configs
• OpenClaw Skills Developer Guide (2026) — Skill architecture, security boundaries, private registry setup
• DigitalOcean: The Complete OpenClaw Skills Guide — Beginner-friendly intro with security review tips

---

🤖 MCP Integrations

MCP (Model Context Protocol, originally by Anthropic) is the dominant architecture for new professional-grade OpenClaw skills. MCP servers run as independent processes, offering better isolation and language-agnosticism.

npm install -g mcporter
openclaw mcp add --url https://mcp.github.com/sse --name github
openclaw mcp add --local /path/to/my-mcp-server --name local-tools

Notable MCP Servers

Browse 13,000+ MCP servers at mcp.so or the official MCP server directory.

---

🖥️ UIs & Companion Apps

Web UIs

Desktop Apps

Mobile

---

🌐 Platform Channels & Messaging

Multi-channel tips:

• Each channel has its own conversation namespace — skills can be scoped per channel
• Dedicate a cheap Telegram bot to low-priority tasks; keep WhatsApp for high-priority work
• Channel-level permission sets: restrict dangerous skills to trusted channels only

---

☁️ Deployment & Infrastructure

Managed Hosting

Self-Hosted: Docker Compose

version: '3.8'
services:
  openclaw:
    image: openclaw/openclaw:stable
    restart: unless-stopped
    ports:
      - "127.0.0.1:18789:18789"   # NEVER bind to 0.0.0.0 directly
    volumes:
      - ./data:/data
      - ./skills:/skills
    environment:
      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY}
      - OPENCLAW_MEMORY_DIR=/data/memory
      - OPENCLAW_LOG_LEVEL=warn
    user: "1000:1000"
    read_only: true
    tmpfs: ["/tmp"]
    security_opt:
      - no-new-privileges:true
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:18789/health"]
      interval: 30s
      timeout: 10s
      retries: 3

Self-Hosted: NixOS

services.openclaw = {
  enable = true;
  package = pkgs.openclaw;
  settings = { port = 18789; memoryDir = "/var/lib/openclaw/memory"; };
  environmentFile = "/run/secrets/openclaw-env";
};

See nix-openclaw for the community NixOS module.

VPS Price Comparison

Homelab & Edge

---

🏢 Enterprise Deployment

For teams and organizations with compliance, multi-tenancy, and governance requirements.

Authentication & Access Control

Single Sign-On (SSO):

# SAML 2.0 via Nginx auth_request
# OpenClaw itself is auth-agnostic; put auth in your reverse proxy layer
# Community guide: https://docs.openclaw.ai/enterprise/sso

Multi-Tenancy Patterns

OpenClaw does not have native multi-tenancy. Enterprise teams use one of three patterns:

Pattern A — One instance per team/user (Recommended)

User A → Gateway A (port 18789) → Memory A
User B → Gateway B (port 18790) → Memory B

Complete isolation. Simple. Works with Ansible or Docker Compose per user.

Pattern B — Shared instance, per-channel isolation One Gateway, multiple channels, channel-scoped skills and permissions. Lower cost; shared memory is a risk.

Pattern C — ClawTeam managed multi-tenancy ClawTeam handles auth, routing, and billing. Best for teams >10 people who don't want to manage infra.

Compliance & Auditing

Data Residency

By default, your data goes to your chosen LLM provider's API. For organizations with data residency requirements:

• EU: Use Anthropic EU endpoints or Gemini EU-hosted endpoints; deploy OpenClaw on Hetzner Germany
• Air-gap: Run Ollama locally — zero data leaves your network. See Ollama × OpenClaw Integration Guide
• On-premise: NemoClaw for fully isolated deployments with local GPU inference

Disaster Recovery

# Backup your entire OpenClaw state
tar -czf openclaw-backup-$(date +%Y%m%d).tar.gz \
  ~/openclaw-data/memory \
  ~/openclaw-data/skills \
  ~/.openclaw/config.json

# Restore
tar -xzf openclaw-backup-20260315.tar.gz -C ~/openclaw-data/

The cron-backup skill automates this on a schedule and optionally uploads to S3, Backblaze B2, or any S3-compatible provider.

---

🧠 Memory & Knowledge Systems

Built-in Memory Files

External Memory Integrations

Memory Best Practices

1. Keep SOUL.md short and opinionated. 200–300 words. Specific beats generic.
2. Use AGENTS.md for non-negotiable rules. Security policies, content restrictions, tool limits — the agent cannot override it.
3. Prune MEMORY.md monthly. Old facts accumulate. Review and remove stale entries.
4. Separate work and personal contexts using channel-scoped memory with different channel configurations.
5. Never put credentials in memory files. Use environment variables or the secrets manager.

---

🔒 Security

This section exists because no other Awesome OpenClaw list treats security seriously enough. OpenClaw runs with broad system access. Security is not optional.

Known CVEs & Incidents

Essential Security Tools

SecureClaw (Install First)

SecureClaw by Adversa AI:

• 55 audit checks across the full attack surface
• 15 behavioral rules (~1,150 tokens) against prompt injection
• Maps to OWASP ASI Top 10 and MITRE ATLAS

openclaw plugin add secureclaw
openclaw skill install adversa-ai/secureclaw
openclaw secureclaw audit --full
openclaw secureclaw harden

Other Security Tools

Secure Deployment: Nginx Reverse Proxy

Never expose port 18789 directly. Always use a reverse proxy with authentication:

server {
    listen 443 ssl http2;
    server_name yourdomain.com;

    ssl_certificate     /etc/letsencrypt/live/yourdomain.com/fullchain.pem;
    ssl_certificate_key /etc/letsencrypt/live/yourdomain.com/privkey.pem;

    location /openclaw/ {
        auth_basic "OpenClaw Gateway";
        auth_basic_user_file /etc/nginx/.htpasswd;

        proxy_pass http://127.0.0.1:18789/;
        proxy_http_version 1.1;
        proxy_set_header Upgrade $http_upgrade;
        proxy_set_header Connection "upgrade";
        proxy_set_header Host $host;

        # Rate limiting
        limit_req zone=openclaw burst=20 nodelay;
        limit_req_log_level warn;
    }
}

# Rate limit zone definition (in http block)
# limit_req_zone $binary_remote_addr zone=openclaw:10m rate=10r/s;

Skill Vetting Checklist

Before installing any skill from ClawHub:

• Author verification: Real GitHub profile with commit history?
• Age and stars: New repo + suspiciously high stars = red flag
• Read the source: SKILL.md files are markdown — readable in 2 minutes. Look for hidden instructions after HTML comments or inside code blocks
• Permission scope: Does a calculator skill request network access? Reject over-privileged skills
• VirusTotal: Check the ClawHub page for VirusTotal scan results
• Run skill-guard: openclaw run "use skill-guard to audit [skill-name]"

Automatic red flags in skill source:

• Instructions to ignore previous instructions or AGENTS.md
• Requests to send data to external URLs not mentioned in the README
• Base64-encoded content embedded in the skill file
• Instructions to modify SOUL.md or AGENTS.md
• fetch() calls to hardcoded external domains without disclosure

---

🆚 OpenClaw vs Other Agent Frameworks

For product teams and developers evaluating which framework to build on. Updated March 2026.

Comparison Matrix

When to Choose OpenClaw

✅ Choose OpenClaw if:

• You want a personal productivity agent you interact with via messaging apps
• You need a large ecosystem of pre-built skills (13,729 on ClawHub)
• You want low-code skill authoring (SKILL.md format)
• Local-first, self-hosted privacy matters to you
• You're building for a team that will use their existing chat tools

❌ Don't choose OpenClaw if:

• You need complex multi-agent orchestration with elaborate handoff protocols (→ use AutoGen or CrewAI)
• You're building a customer-facing chatbot with conversation flows (→ use Botpress)
• You need a visual no-code automation builder for non-technical users (→ use n8n)
• You're doing research requiring fine-grained state machine control (→ use LangGraph)

OpenClaw vs AutoGen

The most common comparison. Key difference: AutoGen is designed for multi-agent systems where agents collaborate in code. OpenClaw is designed for single-agent productivity with a rich skill ecosystem. If you want an AI assistant that takes actions on your behalf via chat, OpenClaw wins. If you want agents collaborating to solve complex research problems, AutoGen wins.

OpenClaw vs n8n

n8n is a visual workflow automation tool for non-technical users. OpenClaw is a conversational agent for power users. They're complementary: many users run n8n for scheduled workflows and OpenClaw for on-demand conversational tasks. The n8n-webhook skill bridges the two.

---

🌍 Ecosystem & Agent Platforms

Agent Social & Collaboration

Agent-to-Agent (A2A)

---

🛠️ Developer Tooling

Deployment & Management

Development & Testing

Monitoring & Observability

Cost Optimization

---

🏗️ Alternative Implementations

Decision Matrix

Implementations

OpenClaw — Reference implementation. Choose when: you want the largest skill ecosystem and accept the resource tradeoffs as the cost of maturity.

NanoClaw — Python, mandatory container isolation, ~700 lines (auditable in an afternoon). Choose when: security and auditability matter more than skill count. Regulated environments.

ZeroClaw — Rust, 3.4MB binary, <10ms startup, 7.8MB RAM (194× smaller than OC), direct OpenClaw config import, 22+ providers. Choose when: resource efficiency matters, or you want a migration path from OpenClaw.

Moltis — A trusted Rust-native Claw. One binary — sandboxed, secure, auditable. Built-in voice, memory, MCP tools and multi-channel access. Choose when: your team needs a trustworthy enterprise-grade Rust implementation.

PicoClaw — Go, <10MB RAM, 1-second startup, targets $10 RISC-V hardware. Choose when: IoT, home automation, embedded Linux, cheapest possible hardware.

Nanobot — Python, ~4,000 lines, maximum readability, MCP-only tools. Choose when: learning agent internals, prototyping architectures, academic research.

NullClaw — Zig, 1MB RAM, 678KB binary. Choose when: absolute minimal footprint, microcontroller-adjacent.

---

📖 Tutorials, Blogs & Articles

Getting Started

• Official OpenClaw Quickstart
• Deep Dive: Architecture, Code & Ecosystem
• What Are OpenClaw Skills? A 2026 Developer's Guide
• The Ecosystem: ClawHub and Skills
• Top 6 OpenClaw Tools Developers Are Using in 2026

Intermediate & Advanced

• OpenClaw New Features 2026 — Official blog roundup
• Best Skills, Plugins and Automations: The Ultimate Guide
• Maximize Efficiency: Top 10 Plugins for 2026

Security

• CVE-2026-25253 Post-mortem — Official security analysis
• SecureClaw: Dual Stack Security Plugin
• A Security-First Guide to OpenClaw Alternatives

Alternatives & Ecosystem

• Best OpenClaw Variants — Categorized breakdown
• Everything is Agent: Alternatives Deep Dive
• The Claw Craziness Continues

Learning Paths

• Repovive OpenClaw Roadmap — 9-hour structured course, 165 units

---

🎥 Videos & Talks

---

📚 Research & Academic Resources

Key Papers

Academic Projects

• OWASP Agentic Security Initiative — OpenClaw used as reference implementation for ASI Top 10
• MITRE ATLAS — Globally accessible, continuously updated knowledge base of adversarial tactics and techniques against AI systems, based on real-world attack observations from AI red teams and security groups; SecureClaw includes formal MITRE ATLAS mappings
• Nanobot — Best platform for studying agent internals; easy to instrument

---

💼 Use Cases & Showcases

By Domain

Personal Productivity:

• Email → zero: Route all email through OpenClaw; auto-triage, draft for approval, auto-send routine responses
• Calendar management: Natural language scheduling that actually understands context
• Personal CRM: Remember context from all your relationships; briefed before meetings automatically

Software Development:

• PR review assistant: Daily summaries of open PRs across all your repos
• Incident responder: Monitors logs, pages on-call, drafts incident reports
• Dependency auditor: Weekly scan for outdated or vulnerable packages

Business Operations:

• Meeting notetaker: Record, transcribe, summarize, distribute
• CRM intelligence: Monitor email, update HubSpot/Salesforce automatically
• Report generator: Weekly metrics drafted and sent every Monday morning

Research & Data:

• Literature monitor: Daily digest of new arXiv papers matching your interests
• Competitive intelligence: Monitor competitors' GitHub, blog posts, job postings
• Data pipeline: Scrape → clean → analyze → summarize on a schedule

Showcase Collections:

• hesamsheikh/awesome-openclaw-usecases — The definitive use case collection
• LAMBDASOFT-org/awesome-openclaw-ecosystem — Agent-native platform showcases

---

❓ FAQ & Troubleshooting

Installation & Setup

Q: What are the minimum hardware requirements? A: Node.js 20+ for OpenClaw core. Minimum 1GB RAM (2GB recommended for comfortable operation with several skills). Any modern 64-bit OS. For local models via Ollama: 8GB VRAM minimum; 16GB recommended.

Q: What's the difference between openclaw onboard and openclaw start? A: onboard is a one-time interactive wizard that sets up your config, connects your first channel, and installs essential skills. start just starts the gateway with existing config. Always run onboard first.

Q: Do I need a paid API key to get started? A: No. You can run fully locally using Ollama + an open-source model (Llama 4, Qwen3). The experience is less capable but costs nothing. Most users start with a Claude or OpenAI API key for the best out-of-the-box quality.

---

Gateway & Connection Issues

Q: Gateway starts but I can't connect from my messaging app. A: Check these in order:

1. Confirm port 18789 is not blocked: curl http://localhost:18789/health
2. If using Docker, ensure you're binding correctly: 127.0.0.1:18789:18789 (not 0.0.0.0)
3. If accessing remotely, confirm your reverse proxy is correctly configured
4. Check logs: openclaw logs --tail 50

Q: WhatsApp keeps disconnecting every few hours. A: This is a Baileys (WhatsApp Web library) limitation. Solutions:

• Keep the QR-scanned device active and connected to the internet
• Configure auto-reconnect rules in AGENTS.md, or use a deployment script with watchdog support
• Consider switching to Telegram as your primary channel — it's more stable for server deployments

Q: My Telegram bot stopped responding. A: 99% of the time this is a bot token issue or a conflicting webhook. Check:

openclaw channel status telegram
openclaw channel reconnect telegram

---

Performance & Memory

Q: OpenClaw is using 2GB+ of RAM. Is that normal? A: With the full skill ecosystem loaded, yes. Mitigation strategies:

• Use openclaw skill list --active to audit which skills are loaded
• Uninstall skills you don't use: openclaw skill remove [name]
• Consider ZeroClaw (7.8MB RAM) if resource efficiency is critical

Q: Responses are getting slower over long conversations. A: Context window is filling up. Solutions:

• Enable aggressive compaction: openclaw config set context.compaction_threshold 0.7
• Schedule daily resets: openclaw config set context.reset_schedule "0 4 * * *"
• Configure health checks and auto-restart in your docker-compose.yml

Q: Memory is growing unboundedly. MEMORY.md is now 50,000 words. A: The agent will write memories indefinitely if not pruned. Best practices:

• Set a monthly calendar reminder to review and prune MEMORY.md
• Use the memory-manager skill to auto-summarize and compress old entries
• Hard limit: Add to AGENTS.md: "Keep MEMORY.md under 10,000 words. Summarize and compress when approaching that limit."

---

Skills & Security

Q: I installed a skill and now my agent is behaving strangely. A: Likely prompt injection via a malicious or poorly written skill. Steps:

1. Disable the skill: openclaw skill disable [name]
2. Review its source code for hidden instructions
3. Run openclaw secureclaw audit --full
4. Report to ClawHub if it's genuinely malicious

Q: How do I prevent a skill from accessing the internet? A: Add to AGENTS.md:

[skill-name] must never make external network requests except to [approved-domain].

For stronger enforcement, use agent-access-control skill or switch to NanoClaw which offers container-level isolation.

Q: My API key was accidentally written to a memory log. What do I do?

1. Immediately rotate the key at your provider's console
2. Delete the affected log: rm ~/openclaw-data/memory/[date].md
3. Run openclaw secureclaw audit to check for other exposures
4. Add to AGENTS.md: "Never write, log, or repeat API keys, tokens, or credentials under any circumstances."

---

Costs & Models

Q: My API costs are much higher than expected. A: Common causes:

1. Skills with very broad trigger conditions → audit with cost-tracker skill
2. No session compaction → conversations accumulate context indefinitely
3. Wrong model for the task → use ClawRouter to route cheap tasks to cheap models
4. Context window leaks from a poorly written skill loading large documents

Q: Can I set a monthly spending cap? A: Not natively in OpenClaw core, but:

• The cost-tracker skill sends alerts when you hit thresholds
• Set hard limits directly at your API provider (Anthropic, OpenAI all support this)
• ClawRouter can be configured to stop routing to expensive models after a cost limit

---

Updates & Migrations

Q: How do I safely update OpenClaw?

# Backup first
tar -czf ~/openclaw-backup-$(date +%Y%m%d).tar.gz ~/openclaw-data/

# Update
npm update -g openclaw

# Verify
openclaw --version
openclaw health check

Q: How do I migrate from OpenClaw to ZeroClaw? ZeroClaw supports direct config import:

zeroclaw migrate --from ~/.openclaw
# Memory and config files are compatible; most skills work via MCP bridge

---

🤝 Community & Support

Official Channels

Key People to Follow

Newsletters

---

🗺️ Related Awesome Lists

---

🤝 Contributing

Contributions are welcome! This list aims to be opinionated and high-quality, not exhaustive.

Before submitting a PR:

1. Have you actually used or read this resource? No link-farming.
2. Add a description — every entry needs context: what, who, why.
3. Check for duplicates first.
4. Match the existing table/list format.
5. Security resources get priority review.

See CONTRIBUTING.md for the full guide and CODE_OF_CONDUCT.md for community standards.

To report a malicious skill or security issue: Open an Issue with the security label. Do not submit CVEs as PRs.

---