OpenClaw Orange Paper: From Beginner to Expert

A comprehensive guide based on the OpenClaw Orange Paper v1.1.0, covering architecture, deployment options, model selection, cost optimization, security, and the China ecosystem. Perfect for getting started with OpenClaw.

2026-03-1325 min
OpenClawOrange PaperTutorialDeployAI Assistant

What is OpenClaw

OpenClaw is an open-source multi-channel AI agent framework (MIT license) with over 280,000 GitHub stars. It lets you interact with an AI assistant through 20+ messaging platforms including Telegram, WhatsApp, Discord, Lark, and more. Think of it as an "AI gateway" — a service running in the background that connects your chosen chat platform to AI models, letting you talk to Claude, GPT, Gemini, and others right from your phone.

OpenClaw was created by Peter Steinberger, originally as a personal project to use Claude on WhatsApp. After going open-source in October 2025, it exploded in the AI community. As of March 2026, it has 1,075+ contributors, 13,729 community Skills, and a global "claw raising" community.

In simple terms: ChatGPT is an "advisor" (you ask, it answers), while OpenClaw is an "employee" (it proactively executes tasks). OpenClaw can connect to your messaging platforms, manage emails and calendars, operate browsers, execute Shell commands, and your data stays entirely in your own hands.

Why Choose OpenClaw

OpenClaw earned 280K+ Stars in just a few months, making it the fastest-growing open-source AI tool project. The key is that it solves a real pain point: bringing AI assistants from "web chat windows" into the messaging tools you use every day.

Use Your Favorite Chat Apps

No need to switch apps. Chat with AI directly in Telegram, WhatsApp, Discord, Lark, and 20+ other platforms, as naturally as talking to a friend.

Data Sovereignty

Runs on your own server, with conversations never passing through third-party platforms. Self-hosted = you own your data.

Switch Between Models Freely

Supports Claude, GPT, Gemini, DeepSeek, GLM, Kimi, and 200+ models, with automatic fallback chains for graceful degradation.

13,729 Community Skills

The ClawHub skill marketplace covers image generation, code review, Notion sync, translation, and more. Install with a single command.

24/7 Online

Deployed on cloud servers, unaffected by your local computer's power state. Supports cron jobs for automated daily briefings, monitoring, and more.

Flexible Cost Control

Combine fallback chains + free models (GLM Flash / Gemini Flash) to keep monthly costs as low as $2-5.

The Chinese community calls running and maintaining an OpenClaw instance "claw raising" (Claw = claw = lobster), and users call themselves "claw raisers." The greeting has become "Are you raising a claw yet?" — this term has spread from tech circles to mainstream media. On March 6, nearly a thousand people lined up at Tencent Cloud's Shenzhen headquarters to install OpenClaw, with news headlines reading "Everyone's Raising Claws."

Core Architecture

Understanding OpenClaw's architecture helps with configuration and troubleshooting. OpenClaw uses a three-layer architecture: Gateway → Node → Channel.

Gateway (Entry Layer)

The system entry point, responsible for authentication, message routing, and session management. Listens on port 18789 by default. All messages (from Telegram, Discord, etc.) pass through the Gateway before being dispatched to Nodes. As of v2026.3.7, authentication configuration is mandatory.

Node (Execution Layer)

The runtime environment for AI Agents. Nodes manage tool calls, the memory system, and Skill execution. A single instance can run multiple Agents, each with its own model configuration and tool permissions. This is where the actual "thinking" happens.

Channel (Messaging Layer)

Adapters for messaging platforms. Each platform (Telegram, Discord, WhatsApp, etc.) has a corresponding Channel plugin. Channels handle protocol adaptation — sending/receiving messages, processing rich text, and managing group permissions.

Four-Layer Memory System

OpenClaw's memory isn't just simple chat history — it's organized into four layers:

LayerFilePurpose
SOULSOUL.mdThe Agent's "soul" — personality, behavioral guidelines, worldview
TOOLSTOOLS.mdTool usage guide — which tools are available and when to invoke them
USERUSER.mdUser profile — what the Agent knows about you, accumulated automatically
SessionMEMORY.mdSession memory — key information persisted across conversations
Design Philosophy: OpenClaw creator Peter was deeply influenced by the Unix philosophy — configuration files use Markdown instead of databases, and each component maintains a single responsibility. He has reservations about MCP, calling it "unnecessary complexity," and prefers native tool integrations.

Core Features

OpenClaw is more than just a chatbot — it's a full-featured AI agent platform:

Skill Ecosystem (ClawHub)

ClawHub is OpenClaw's official skill marketplace, hosting 13,729 Skills as of March 2026. Skills are loaded in three priority levels: user-defined → project-level → global. Install a new skill with a single command:

Install a Skill
openclaw plugins install <skill-name>

Popular skills include: image generation (DALL-E / Flux), Notion sync, GitHub management, calendar integration, web clipping, translation, code review, and more.

Security Warning: About 20% of Skills on ClawHub have quality issues (spam/duplicate/malicious). The ClawHavoc supply chain attack in February 2026 affected 135,000+ devices, with around 800 malicious Skills attempting to steal user credentials. We recommend only installing highly-starred Skills, reviewing source code before installation, and using the awesome-openclaw-skills curated list.

Automated Workflows

  • Cron Jobs: daily briefings, monitoring alerts, scheduled backups
  • Webhook Triggers: integrate with GitHub, Notion, Lark, and other systems
  • Event-Driven (Hooks): automatically execute specific actions when messages arrive
cron-config.json
// openclaw.json cron job example
{
  "cron": [
    {
      "schedule": "0 8 * * *",
      "prompt": "Summarize today's tech news, limit to 5 items"
    }
  ]
}

Multi-Agent Orchestration

OpenClaw supports configuring multiple Agents, each with its own model, personality, and tool permissions. By defining personalities through SOUL.md and pairing them with different models, you can have one assistant handle coding, another for writing, and another for data analysis. Agents can even communicate and collaborate with each other.

Browser & Tool Calling

OpenClaw has built-in browser automation capabilities, enabling web browsing, form filling, and data extraction. Combined with Shell command execution, file I/O, and other tools, Agents can accomplish complex tasks far beyond the scope of chatting — a true "digital employee."

Supported Platforms & Models

International Platforms

TelegramDiscordWhatsAppSlackiMessageSignalMattermostTeamsMatrix

Chinese Platforms (via openclaw-china plugin)

DingTalkQQ BotWeComLarkWeChat (indirect)

Connect in three steps via openclaw plugins install @openclaw-china/channels

AI Model Support

ClaudeGPTGeminiDeepSeekGLM (Zhipu)KimiQwen (Tongyi)DoubaoMiniMaxBaidu ERNIEOllama (Local)LM Studio (Local)200+ More

Deployment Options Comparison

OpenClaw supports multiple deployment methods. The Orange Book compares 10+ options in detail. Here's a summary of the main choices:

Deploy MethodDifficultyTimeCostMaintenance
Clawfy One-Click DeployRecommendedZero barrier~30 secPay-as-you-goFully managed
Alibaba / Tencent Cloud One-ClickLow10-20 minServer + APISemi-managed
Docker DeployMedium30-60 minServer + APISelf-maintained
Local npm InstallLow10-20 minAPI onlySelf-maintained
1Panel Dashboard DeployLow10-15 minServer + APIPanel-managed
For users in China: Alibaba Cloud lightweight server one-click deploy is the most popular option domestically (starting at ~68 CNY/year for new users). Paired with the openclaw-china plugin for DingTalk/QQ + DeepSeek or GLM-5 as the primary model, the total monthly cost can be kept under 100 CNY.

Deploy with Clawfy in One Click

Clawfy is a managed hosting platform built specifically for OpenClaw. Compared to self-hosting, Clawfy handles all the tedious steps for you — server purchase, Docker installation, image pulling, and environment configuration:

No BYOK Required — API access is unified through OpenRouter. Just purchase Credits and you can use all models
Snapshot-Warmed Pool — Servers are pre-created, with deployment taking about 30 seconds (traditional methods take 30-60 minutes)
Pay-as-You-Go — Unlike SimpleClaw's fixed $49/month fee, Clawfy charges only for what you use, with a lower starting cost
Fully Managed — Server monitoring, billing management, and one-click start/stop are all handled from the dashboard

Quick Start Guide

Deploying OpenClaw with Clawfy takes just three steps. The entire process takes under 5 minutes, with no technical background required.

Step 1: Create a Telegram Bot

1

Find @BotFather on Telegram

Open Telegram and search for @BotFather, Telegram's official Bot management tool.

2

Create a New Bot

Send /newbot, then follow the prompts to set the Bot's name and username. Once complete, you'll receive a Bot Token (formatted like 123456789:ABCdefGHI...).

3

Copy the Token

Save this Token — you'll need it in the next step for deployment.

Step 2: One-Click Deploy

1

Sign in to Clawfy

Visit clawfy.io and sign in quickly with your Google account.

2

Choose Your Configuration

Select an AI model (Gemini Flash is recommended for beginners — affordable and fast), choose your chat channel (Telegram), and paste your Bot Token.

3

Click Deploy

Click the "Deploy" button. Clawfy will automatically provision a dedicated server, install, and configure OpenClaw for you. The entire process takes about 30 seconds.

Step 3: Pair and Start Using

1

Send /start

Find the Bot you just created in Telegram and send /start. The Bot will return a pairing request.

2

Approve Pairing in Clawfy

Go back to the Clawfy dashboard and click "Approve" on the instance detail page to confirm the pairing.

3

Start Chatting!

Once paired, you can freely chat with your AI assistant in Telegram. Try asking it anything!

Example Conversation
You: Summarize today's top HackerNews articles
AI: Browsing HackerNews...

Here are the top 5 trending articles today:
1. "Rust Compiler Self-Bootstrapping Complete" — The Rust team announced...
2. "GPT-5.4 Released" — OpenAI launches its latest model...
3. "YC W26 Demo Day Highlights" — The most-watched startups this season...
4. "Quantum-Safe Encryption Standard Takes Effect" — NIST publishes...
5. "SpaceX 20th Recovery Landing" — Setting a new record...

Model Selection Guide

The Orange Book provides a detailed comparison of 14+ models by price and characteristics. Here's a quick reference table sorted by input price (price per million tokens):

ModelInput /1MOutput /1MOne-Line Summary
Ollama / LM StudioFreeFreeUses local compute only
GLM Flash / ERNIE SpeedFreeFreeCloud-based free tier
Doubao 1.5 Lite-32k$0.042Cheapest cloud conversations
DeepSeek-V3.2$0.14$0.28Best value for money
Qwen 3.5 Coder$0.22$1.00Best value for code
Gemini 3 Flash$0.50$3.00Affordable, strong multimodal
Kimi K2.5$0.60$3.00Chinese language flagship
GLM-5$0.80$2.56Top Chinese code model
Claude Haiku 4.5$1.00$5.00Lightweight international
Gemini 3 Pro$2.00$12.00Google flagship
GPT-5.4$2.50$15.00OpenAI flagship
Claude Sonnet 4.6$3.00$15.00Best for agentic tasks
Claude Opus 4.6$5.00$25.00Most capable, most expensive

Five Recommended Configurations

Option 1: Ultra Budget (avg <$5/mo)

  • Primary: DeepSeek-V3.2 ($0.14/$0.28)
  • Fallback: Qwen 3.5 Plus ($0.40/$1.20)
  • Heartbeat/Cron: GLM-4.5-Flash (Free)
  • Reasoning: DeepSeek-R1 ($0.55/$2.19)

Ideal for individual developers and exploration

Option 2: Chinese Model Value (avg $5-15/mo)

  • Primary: GLM-5 ($0.80/$2.56)
  • Fallback: DeepSeek-V3.2 ($0.14/$0.28)
  • Reasoning Boost: Kimi K2.5 ($0.60/$3.00)
  • Simple Tasks: GLM-4.5-Flash (Free)

Best for Chinese-language users prioritizing stability

Option 3: International Balanced (avg $10-30/mo)

  • Primary: Claude Sonnet 4.6 ($3.00/$15.00)
  • Lightweight: Claude Haiku 4.5 or Gemini Flash
  • Complex Tasks: Claude Opus 4.6 (on demand)
  • Heartbeat/Cron: Gemini Flash (free tier)

For optimal Agent performance with sufficient budget

Option 4: Optimal Hybrid (avg $5-20/mo, Recommended)

  • Complex Tasks: Claude Sonnet 4.6
  • Daily Chat: DeepSeek-V3.2
  • Heartbeat/Cron: Gemini Flash or local Ollama
  • Fallback Chain: Sonnet → Haiku → DeepSeek

Best choice for most users, balancing quality and cost

Clawfy User Tip: On Clawfy, you access 290+ models through OpenRouter without needing to register API Keys separately. Just purchase Credits and switch between any model at any time. Use Gemini Flash for daily chat (nearly free), and switch to Claude Sonnet/Opus when you need deep analysis.

Cost Optimization Strategies

The Orange Book specifically warns: API costs are the biggest expense of running OpenClaw. Horror stories of "waking up to a $1,100 API bill" are common in the community. Here's why:

  • Each Agent "thought" involves multi-step reasoning — a simple task may trigger 5-10 API calls
  • Skill descriptions are injected into the system prompt, increasing input tokens per request
  • The memory system (MEMORY.md + Daily Logs) attaches context to every request
  • Agents run 24/7, with cron jobs continuously triggering API calls
  • Multi-step reasoning + multi-tool calls can consume tens or even hundreds of times more tokens than traditional chat

Core Cost-Saving Strategy: Fallback Chains

Fallback chains are OpenClaw's most important cost-saving mechanism. When the primary model is unavailable, it automatically degrades to a cheaper model — but the smarter approach is to proactively use this to control costs:

Fallback Configuration Example
{
  "agents": {
    "defaults": {
      "model": {
        "primary": "anthropic/claude-sonnet-4-6",
        "fallbacks": [
          "anthropic/claude-haiku-4-5",
          "deepseek/deepseek-chat"
        ]
      }
    }
  }
}
StrategyRelative Cost
All Claude Sonnet100% (baseline)
Sonnet + Haiku Fallback~50-60%
Sonnet → Haiku → DeepSeek (three-tier)~5-20%
Pure DeepSeek-V3~5%
Local Ollama0% (electricity only)

Setting Budget Limits

We strongly recommend all users set a daily budget cap — even if money isn't an issue, a $5 daily limit can protect your wallet when an Agent enters a reasoning loop:

Budget Limit Configuration
{
  "agents": {
    "defaults": {
      "budget": {
        "maxTokensPerDay": 500000,
        "maxCostPerDay": 5.00
      }
    }
  }
}
Core Cost Control Principle: Use the right model for the right task. Simple Q&A uses $0.14/M DeepSeek-V3, complex reasoning uses $3.00/M Claude Sonnet, and free Gemini Flash or Ollama handles heartbeat and cron tasks. A three-tier fallback chain + daily budget cap — these two strategies alone can cut monthly costs from triple digits to double or even single digits.

Security Advisory

OpenClaw's security model is built on "trust nothing by default," but creator Peter himself admits: "This is all vibe code. Prompt injection hasn't been solved. There are absolute risks." In less than 5 months of history, OpenClaw has already experienced at least 7 major security incidents.

CVE-2026-25253: Remote Code Execution Vulnerability

CVSS 8.8/10 (High severity). WebSocket origin header bypass allowed attackers to execute arbitrary code on OpenClaw instances. Fixed in v2026.3.2.

ClawHavoc Supply Chain Attack

135,000+ devices affected. About 20% of Skills on ClawHub were confirmed malicious at peak. We recommend only installing highly-starred Skills and using the awesome-openclaw-skills curated list.

30,000+ Unauthenticated Exposed Instances

Security researchers discovered over 30,000 OpenClaw instances exposed on the public internet with no authentication configured. As of v2026.3.7, Gateway authentication is now mandatory. Make sure to check that your instance has authentication configured.

Anthropic OAuth Ban / Google Account Suspensions

In January 2026, Anthropic banned Claude Pro/Max subscription accounts from connecting to OpenClaw via OAuth. In early February, Google mass-suspended OpenClaw users' Google accounts (including Gmail, Drive, and all other services). Using Anthropic requires API Keys, and Google-related Skills should use a dedicated Workspace account.

Security Recommendations: (1) Keep OpenClaw updated to the latest version; (2) Configure Gateway authentication (token or password); (3) Review source code before installing Skills; (4) Set spending limits; (5) Do not visit untrusted websites on devices running OpenClaw (ClawJacked zero-click vulnerability). If you use Clawfy, these security configurations are handled by default.

China Ecosystem

The "cloud claw raising" community has over 100,000 users, government support policies have been issued, and OpenClaw's adoption in China has exceeded all expectations. On March 8, 2026, the Shenzhen Longgang AI (Robotics) Bureau published a draft support policy for OpenClaw usage, and on March 6, nearly a thousand people lined up at Tencent Cloud headquarters to install it.

Chinese Claw Products Overview

Major Chinese tech companies have launched their own "lobster" products. As of March 2026, there are 10+ options available:

ProductCompanyDefault ModelKey Selling Point
MaxClawMiniMaxM2.518-sec deploy, lowest price
AutoClawZhipu AIGLM-596 preset Skills, free
QClawTencentKimi-2.5WeChat/QQ direct connect
ArkClawByteDance/VolcanoSeed 2.0Deep Lark integration
WorkBuddyTencentHunyuan/GLM etc.Free, WeCom/QQ integration
LobsterAINetEase YoudaoMulti-modelGUI interface, sandboxed
CoPawAlibaba TongyiQwen/OllamaEdge + cloud dual deploy
miclawXiaomiMiMoNative mobile, Mi Home IoT
Buying Tip: Most packaged products lock in a default model and don't allow free model switching like the original OpenClaw. If model flexibility is important to you, consider the original OpenClaw or products that support multiple models (AutoClaw, WorkBuddy, LobsterAI). International users are recommended to use Clawfy for one-click deployment of the original OpenClaw for maximum flexibility.

Frequently Asked Questions

Is OpenClaw free?

OpenClaw itself is MIT open-source and free. However, running it requires two costs: a server (local computer or cloud server) and AI model API fees. If you use local models (Ollama), API costs can also be zero. In short: the software is free, but compute is not.

How is OpenClaw different from ChatGPT?

ChatGPT is an "advisor" (you ask, it answers), while OpenClaw is an "employee" (it proactively executes tasks). OpenClaw can connect to your messaging platforms, manage emails and calendars, operate browsers, execute Shell commands, and your data stays entirely in your own hands. The tradeoff is that you need to deploy and maintain it yourself.

I'm not technical. Can I use Clawfy?

Absolutely! Clawfy is designed specifically for non-technical users. All you need is a Google account and Telegram, and you can deploy in 30 seconds. All server management, Docker configuration, and environment setup is handled automatically by Clawfy.

How much does it cost per month?

It depends on your usage patterns and model choices. Reference ranges: completely free (local models) → $2-5/mo (DeepSeek primary) → $5-15/mo (GLM-5 primary) → $10-30/mo (Claude Sonnet primary). The biggest cost trap is that OpenClaw's multi-step tool calls consume large amounts of tokens — always set spending limits.

Is it secure? Will my data leak?

OpenClaw is self-hosted, with data stored on your own server by default. However, be aware of three security risks: (1) CVE-2026-25253 RCE vulnerability (fixed, make sure to update); (2) ClawHavoc supply chain attack (review source code before installing third-party Skills); (3) Gateway exposed to the public internet needs authentication configured.

Can I use Chinese models? How do they perform?

Absolutely. DeepSeek-V3 ($0.14/M input) and GLM-5 ($0.80/M input) are the most popular choices among Chinese users. GLM-5's coding ability approaches Claude Opus 4.5 level, while DeepSeek-V3 offers extreme value. We recommend mixing them with the Fallback mechanism.

How is Clawfy different from SimpleClaw?

SimpleClaw uses a flat monthly fee ($49/mo) + BYOK model, where you need to register and manage your own API Keys. Clawfy uses pay-as-you-go Credits — no API Key management needed, lower starting cost, and you only pay for what you use.

Can OpenClaw and Claude Code be used together?

Yes, and it's actually a recommended setup. The community has developed openclaw-claude-code-skill, which bridges the two via the MCP protocol. OpenClaw handles messaging platform integration and life automation, while Claude Code handles programming tasks. Together, they form the most complete AI-driven workflow of 2026.

Can it connect to WeChat?

Yes, but not directly to personal WeChat. Through the openclaw-china plugin's WeCom (enterprise WeChat) self-built app approach, you can indirectly connect to personal WeChat. DingTalk and QQ Bot are the easiest to set up. Lark is one of the most popular integration channels in China.

Related Resources

This guide is based on the OpenClaw Orange Book v1.2.0 (2026-03-12) written by Huashu (Bilibili: AI Evolution - Huasheng), edited and adapted for the Clawfy platform.

Deploy your AI assistant with Clawfy

No SSH, Docker, or CLI. Pick a model, connect your Telegram Bot, and go live in one click. Pay as you go, stop anytime.

Get Started Free