OpenClaw Orange Book - A complete one-stop guide from beginner to expert (PDF file)

"OpenClaw Orange Book" is by Hua Shu (Bilibili/YouTube: AIThe OpenClaw Comprehensive Reference Manual, written by Evolutionary Theory (Peanut), covers everything from architectural principles to deployment schemes, from model configuration to security costs. The *OpenClaw Orange Book* is a comprehensive resource for OpenClaw "shrimp farmers."EssentialGuide, in-depth analysis of thisopen sourceThe project's three-tier architecture, four-tier memory system, and self-expansionAgentThe design philosophy encompasses both on-premises and cloud vendors.One-clickDeployment, practical experience with 20+ channel access, security pitfalls in the ClawHub skills ecosystem, and more.ClaudeCost control strategies for domestically produced models.

freeGet theOpenClaw Orange BookThe original PDF report is available by scanning the QR code and replying with the following code: OpenClaw Orange Book

Understanding OpenClaw

What is OpenClaw?

OpenClaw isopen sourceSelf-managedAI AgentThe system allowsAIFrom traditional chat tools to digital employees capable of autonomous tasks. (If you've used Chat...)GPTYou'll know that it's essentially a question-and-answer model—you ask a question, it answers. But OpenClaw is completely different; it's a...AI AgentThe platform can connect to more than 20 messaging channels (including WhatsApp, Telegram, Lark, DingTalk, Discord, etc.) to proactively perform tasks, manage schedules, process emails, operate browsers, and call up various tools.

With ChatGPTThe core differenceChatGPTThe interaction mode is passive response, running in a webpage or app, and the data is stored in Open.AIThe server only supports...GPTSeries models, and notopen sourceYes. OpenClaw supports autonomous task execution, running on a self-hosted server, with complete local data control, and supports...Claude,GPTDeepSeekGeminiEven the local Ollam model, fully licensed under the MIT Licenseopen source.

A Brief History of OpenClaw

• November 2025: ClawdBot is launched (weekend project)

• Mid-January 2026: 60,000 Stars gained in 72 hours

• January 27, 2026: Due to the name andClaudeSimilarly, renamed Moltbot

• January 30, 2026: Renamed OpenClaw again.

• Early February 2026: CVE-2026-25253 RCE vulnerability exposed, 50,000+ instances vulnerable to attack; ClawHavoc supply chain attack outbreak.

• February 14, 2026: Founder Peter Steinberger joins OpenAI

• March 3, 2026: Reached the top of GitHub (over 250,000 stars)

• March 8-9, 2026: The Ministry of Industry and Information Technology issued a security risk warning for OpenClaw.

"Shrimp farming" cultural phenomenon

Because the mascot is a lobster, the Chinese community refers to running OpenClaw as "raising lobsters," and users call themselves "lobster farmers." The greeting has become "Have you raised lobsters yet?" These interesting cultural tags lower the barrier to dissemination, giving a technological project the attributes of social currency.

On March 6, 2026, nearly a thousand people lined up at Tencent Cloud headquarters in Shenzhen to experience the installation of OpenClaw. On March 8, in Longgang District, Shenzhen...AIThe (Robotics) Bureau has released a draft for public comment on support measures for the use of OpenClaw.open sourceIt is rare in the country for a project to attract the policy attention of local governments.

The OpenClaw ecosystem has also spawned Moltbook—a dedicated platform for...AI AgentThe social media platforms used. As of the end of February 2026, there were 32,912 registered users.AI AgentIt has 2,364 sub-communities, 3,130 posts, and 22,046 comments. Thousands of OpenClaw instances post, comment, and discuss philosophical questions on it. This is probably...AI AgentThe first large-scale experimental field for moving from "tools" to "social existence".

OpenClaw's technical architecture

Three-tier architecture design

OpenClaw adopts a three-layer architecture of Gateway-Node-Channel, using WebSocket as the communication bus to decouple the control plane, device execution, and message channels.

• GatewayIt is the central control plane, which maintains the WebSocket service, manages sessions, and schedules...AgentIt defaults to binding to ws://127.0.0.1:18789, and only one Gateway instance runs on each host. This is because channels like WhatsApp Web require exclusive sessions, and multiple instances would cause login conflicts.
• NodeIt is the device-side execution node, responsible for local operations, including camera, screen recording, system.run, etc.
• ChannelIt is a messaging channel access layer that connects more than 20 instant messaging platforms, including WhatsApp, Telegram, Discord, Slack, Lark, DingTalk, etc.

The gateway uses a loopback-first design, binding to only localhost (127.0.0.1) by default, with all traffic looping back locally. This means no external ports are exposed, ensuring inherent security. Nodes on the same machine connect directly to the gateway via WebSocket. When remote access is needed, it is exposed through the Tailscale Serve/Funnel, without directly exposing ports.

Four-layer memory system

Memory is the core capability that distinguishes OpenClaw from ordinary chatbots. Four layers of memory, from an immutable identity kernel to real-time dialogue, construct complete contextual continuity.

• The SOUL layer is stored in the SOUL.md file.It is permanent and immutable.AgentPersonality, values, and core identity definitions should not be modified after creation.
• The TOOLS layer contains Skills and Extensions.Load on demand is a list of currently available tools and skills that changes dynamically as they are installed and loaded.
• The USER layer is stored in MEMORY.md and the vector database.It is a persistent set of user preferences, decisions, and historical facts that supports semantic search.
• The session layer is stored in memory and in sessions.json.It is the current real-time context of the conversation at the session level, and it will be compressed when the token is exhausted.

automaticMemory preservation mechanism

When a session approaches its token limit (default threshold is approximately 4000 tokens), OpenClaw triggers a silent agentic turn.AgentRound): First, detect that the token usage is close to the limit, and trigger the Pre-Compaction process;AgentA hidden turn is executed in the background, writing important memories to MEMORY.md and the Daily Log; finally, the context is compressed, old messages are compressed or truncated, and token space is freed up. This process is not visible to the user (it returns NO_REPLY).

This mechanism ensures that even in extremely long conversations, key information is not lost as the context window is swiped.Claude The context of tools like Code disappears after the session ends, but OpenClaw achieves true persistent memory through the file system.

Agentwork area

• AGENTS.md:AgentThe definition of identity, behavioral boundaries, and response style are essentially a documented version of the system prompt, loaded every time a session starts.
• SOUL.mdThe immutable core of personality, definedAgent"Who is it?" is loaded every time a session starts.
• USER.mdStructured information about the user (name, preferences, relationships) is loaded when the Main session starts.
• MEMORY.mdLong-term memoryAgentPersistent facts and decisions actively written during the conversation are only loaded in the main session.
• HEARTBEAT.mdDefine scheduled tasks and proactive behaviors (such as checking task status every 30 minutes), which are loaded when Gateway starts.
• memory/Daily Logs directory, sorted by dateautomaticCreate, append-only write, read today's + yesterday's logs
• skills/Workspace-level skills, highest priority (higher than global and built-in skills), scanned at session startup.
• sessions.jsonSession metadata storage records the state and history of each session, and allows for on-demand retrieval.

Design Philosophy

Anti-MCP stance

Deployment Overview

• Local installationSuitable for developers and users who want complete control over their data; requires Node.js version 22+ and is compatible with macOS/Linux.recommend,completelyfreeIt does not have a built-in model and is relatively easy to use.
• Docker deploymentSuitable for scenarios requiring environment isolation, easy migration, or long-term operation on servers. Requires Docker Engine. It is the preferred choice in China and is beginner-friendly. It does not have built-in models and has a moderate level of difficulty.
• domestic cloud vendorsOne-clickdeployIt is the preferred choice for most domestic users and is supported by all mainstream cloud vendors. The main differences lie in pricing strategies and integration with the IM ecosystem.
• Alibaba CloudThe platform with the richest community resources in China, pre-installed with an image, ready to use out of the box. Configuration includes 2 vCPUs + 2 GiB RAM + 40 GiB ESSD system disk. Limited-time flash sale price: 9.9 RMB/month; regular annual subscription discount as low as 68 RMB/year. Includes the qwen3.5-plus model by default; the first month of the Bailian Coding Plan Lite is 10 RMB (18,000 uses/month). Supports IM channels such as DingTalk and Lark.
• Tencent CloudIt fully supports four major IM tools, and the Coding Plan model package offers excellent value for money.recommend2-core 4G configuration (ideal configuration), minimum 2-core 2G can run. New user package 2-core 4G is about 17 yuan/month, starting from 99 yuan/year. Coding Plan starts from 7.9 yuan for the first month, including HY 2.0 Instruct, GLM-5, kimi-k2.5, MiniMax-M2.5 and other models. Supports full coverage of four major IM: WeChat Work, QQ, DingTalk and Lark, and supports "limited-time same price renewal" promotion.
• BaiduintelligentcloudThe trial cost is the lowest, with a first-month trial cost of 0.01 yuan (limited to 500 units per day).recommend2-core, 4GB RAM, 4Mbps bandwidth configuration, typically priced between 70-140 RMB/month. Integrates Qianfan platform models from the Wenxin, Qwen, and DeepSeek series. A unique feature is its exclusive Baidu search/Baidu Encyclopedia capabilities; 7 official Qianfan skills are now available on ClawHub.
• Huawei CloudFlexus L instances offer the strongest enterprise-level security and compliance capabilities, making them suitable for existing enterprise users within the Huawei ecosystem. Prices range from approximately 85-155 RMB per month, with no particularly attractive new user discounts. Activation must be performed separately in the MaaS console.AIThe model requires at least 5 deployment steps (creating an instance → EIP → security group → installation → configuring the model). Its advantages include enterprise-level security compliance and support for...automaticIt offers expanded and richer MaaS models.
• Volcano EngineLark's deep integration and the server + model combination package of 19.8 yuan/month is currently the most cost-effective solution.recommendDual-core, 4GB RAM, supports both cloud server and cloud phone deployment. Promotional price: ¥9.9/month; Ark Coding Plan bundled package: ¥19.8/month (server + model). The Ark platform offers a rich set of built-in, usable models. Supports Lark (deeply integrated), WeChat Work, DingTalk, and QQ.
• Coze Code (Button Programming)Zero-barrier solution: No server, no coding, no environment configuration required; deployment complete in 1 minute. A subscription is required to deploy OpenClaw: ¥49/month (basic) or ¥99/month (advanced).freeVersion not supported. Multiple built-in models are available: Seed 2.0, DeepSeek, GLM-4.7, etc., and third-party APIs can be integrated. A key feature is that all models, online search, and raw image skills are pre-configured, and Kouzi programming skills can be directly loaded.

Initial configuration critical

• v2026.3.7 introduced Breaking Change.Gateway authentication now requires explicit setting of `gateway.auth.mode`. Failure to do so will prevent the Gateway from starting. This is to address a security vulnerability previously exposed to over 30,000 unauthenticated instances on the internet.
• The authentication mode must be set in the configuration file located in the ~/.openclaw/workspace directory.Choose token authentication (recommendFor API integration) or password authentication (recommend(For web UI access).
• Model costs are the biggest expense.Server costs have generally dropped to very low levels (9.9-99 RMB/year); the real ongoing cost lies in model usage. When choosing a platform, focus on the model package price, not just the server price.

Unified access process

OpenClaw connects more than 20 chat platforms through a unified Gateway architecture. All channels share the same three-step access mode: create credentials on the platform → write to openclaw.yaml → start Gateway → complete pairing.

Multiple channels can run simultaneously, messagesautomaticRouting to the corresponding platform. Pairing mode (dmPolicy: pairing) is enabled by default; unknown senders require a verification code to communicate with the bot.

International Platform

• TelegramIt is the official OpenClaw website.recommendThis is an easy-to-use introduction method that uses a long-polling pattern, where the bot actively polls the Telegram server to retrieve messages, requiring no public IP address, reverse proxy, or port forwarding. It works correctly in local development, behind NAT, and within a firewall. Deployment can be completed in 5 minutes with zero learning curve.
• DiscordSuitable for community management and team collaboration scenarios, it requires creating an Application and a Bot in the Developer Portal. The permission settings involve a few steps, but the documentation is complete, and it takes 15-20 minutes.
• WhatsAppThis is the most popular channel in the OpenClaw community. It uses the Baileys library to connect via QR code scanning, does not require the WhatsApp Business API, and takes 10-15 minutes. It is recommended to use a separate number, not your main number.
• SlackSuitable for internal use by enterprises and teams, it requires creating an app on the Slack API platform and configuring multiple permissions. It uses Socket Mode (WebSocket) by default, does not require a public URL, and takes 25-40 minutes.
• SignalProvides end-to-end encrypted communication, connecting to the Signal network via the Signal-CLI tool, which takes 20-30 minutes.
• iMessageAccess is achieved through BlueBubbles bridging, which requires a constantly running Mac as a BlueBubbles Server and takes 30-45 minutes.

Domestic platforms

• QQIt is the most accessible OpenClaw service for domestic users.SimpleThe official Tencent QQ Bot capability has been made available to OpenClaw; binding can be completed in just one minute by scanning a QR code. It supports multimedia messages such as Markdown, images, voice messages, and files, and can be used on both mobile QQ and desktop QQ.
• FeishuNative built-in support has been provided since OpenClaw 2026.2. It uses WebSocket event subscription to support private chat, group chat, multimedia messages such as photos/files/videos, and takes 15-20 minutes to process.
• DingTalkAccessing OpenClaw via a community plugin uses Stream mode (WebSocket long connection) for message reception, requiring no public IP address and taking 20-30 minutes. DingTalk has not yet received official built-in support from OpenClaw, but the community solution is already very mature.
• WeChat for BusinessThere are two access modes:AgentThe two modes (classic XML callback mode and Bot mode, JSON callback mode with native stream support) have been adopted and validated by public cloud platforms such as Tencent Cloud, Volcano Engine, and China Telecom Cloud.
• WeChatPersonal WeChat accounts have the greatest demand but are also the most complex. There is no official bot API for personal WeChat accounts; all solutions are unofficial, and the risk of account suspension always exists.recommendThe solution uses WeChat Work as an intermediary, which is legal and compliant. Within the WeChat ecosystem, it requires access to the WeChat Work management backend.

How Skills Work

• Workspace-level Skills,lie in.
• User-level SkillsLocated in the ~/.openclaw/skills/ directory, it has global effects. Skills installed via ClawHub or placed manually are located here.
• Built-in SkillsThere are 55 in total, released with the OpenClaw version. No installation is required; they are ready to use out of the box.

ClawHub ecosystem

PopularSkillsrecommend

• Top 10 Must-Have ItemsGmail/Google (email sending and receiving, calendar management, Google Docs reading and writing)Agent BrowserautomaticSummarize (videos, web pages, email content)automatic(Summary), GitHub (repository management, issue handling, PR review),Claude Code (bridged via MCP protocol)Claude Code capabilities), Web Search (online search), File Manager (local file operations), Calendar (schedule management), Translator (multilingual translation), Image Generator (…AIImage generation.
• practicalsuggestionDon't install too many skills at once. Each skill increases the length of the system prompt and occupies the context window. It's recommended to start with 3-5 skills from the Top 10 that you truly need, and then gradually expand the list once you're familiar with them.

Skills Security Warning

Model Provider Overview

• Built-in ProviderIncluding Anthropic, OpenAIIt can be used with Google, Zhipu (zai), etc., without any additional configuration; simply set an API Key.
• Custom ProviderApps such as DeepSeek, Doubao, and Kimi need to be manually added in models.providers.
• Fallback mechanismThe core cost-saving strategy is: when the main model is unavailable.automaticSwitch to the alternative.

International Model

• Anthropic ClaudeIt is the default model provider for OpenClaw and is recognized by the community.AgentThe model with the best task performance.Claude Sonnet 4.6 significantly outperforms other models in terms of accuracy and stability when using tools, with an input price of $3.00/1M tokens, an output price of $15.00/1M tokens, and a context window of 200K.
  • NoteAnthropic has blocked OAuth authentication. (Use...)Claude Pro/Max subscription accounts that connect to OpenClaw via OAuth may receive warnings or even have their accounts locked. Currently, the only legitimate path is to use an API Key (pay-as-you-go).
  • Money-saving tipsBatch API is available with a 50% discount (both input and output are half price).Prompt Caching can reduce the cost of repetitive contexts by up to 90%.

  Domestic Models

  • DeepSeekIt boasts unparalleled cost-effectiveness. DeepSeek-V3.2.2 has an input cost of only $0.14/1M tokens and an output cost of $0.28/1M tokens, making it the most commonly used low-cost model in the OpenClaw community. However, DeepSeek occasionally experiences delays or even becomes unavailable during peak periods, so it is not recommended as the sole provider and must be paired with a fallback model as a backup.
  • GLMIt is the choice with the strongest coding capabilities among domestically produced models. The GLM-5 achieved a score on the SWE-bench.open sourceThe model achieved the highest score with a price of only $0.80/1M input and $2.56/1M output. Even better, OpenClaw has a built-in zai provider, making configuration extremely simple.SimpleGLM-4.5-Flash and GLM-4.7-Flash are completely compatible.freeIdeal for heart-pounding tasks andSimpledialogue.
  • A Thousand Questions on General Principles Version 3.5 is the version Alibaba will release in February 2026.up to dateVersion (397B total parameters/17B activation, MoE architecture, already)open sourceQwen 3.5 Coder has an input cost of only $0.22/1M tokens and an output cost of $1.00/1M tokens, making it extremely cost-effective for code users.
  • Doubao Seed 2.0 Pro has an input price of $0.47/1M tokens and an output price of $2.37/1M tokens. It is a flagship inference model, comparable to...GPT-5.2. Doubao 1.5 Lite - 32k input costs only $0.042/1M tokens, making it one of the cheapest options currently available.
  • Kimi K2.5 input price $0.60/1M tokens, output $3.00/1M tokens, is the Dark Side of the Moon.up to dateFlagship, ChineseAgentIt performs well in long context scenarios.
  • MiniMax M2.5 (230B parameters) scored 80.2% on SWE-Bench, demonstrating outstanding coding ability. Its input price is $0.50/1M tokens, and its output price is $2.00/1M tokens.

  Five setsrecommendConfiguration Scheme

  • Option 1: Extreme cost savings (average monthly < $5) The primary device is DeepSeek-V3.2 ($0.14/$0.28), with Qwen 3.5 Plus ($0.40/$1.20) as an alternative. For heart rate/cron monitoring, GLM-4.5-Flash is used.freeFor inference tasks, use DeepSeek-R1 ($0.55/$2.19). Suitable for individual developers and for learning and exploration. The risk is that DeepSeek may experience delays during peak periods, requiring a fallback mechanism.
  • Option 2: Domestic products with good value (average monthly cost of $5-15) The primary system uses GLM-5 ($0.80/$2.56), with DeepSeek-V3.2 as an alternative ($0.14/$0.28) and Kimi K2.5 ($0.60/$3.00) for enhanced inference.SimpleThe task uses GLM-4.5-Flash (freeSuitable for domestic users who prioritize a Chinese user experience and stability. GLM-5 boasts strong code capabilities and low latency.
  • Option 3: International Balance (Monthly average $10-30) Main useClaude Sonnet 4.6 ($3.00/$15.00), LightweightClaude Haiku 4.5 orGemini Flash, for complex tasksClaude Opus 4.6 (upgrade as needed), for heartbeat/Cron.Gemini Flash (free(Limit). Suitable for those seeking...AgentFor users with the best results and sufficient budget.ClaudeexistAgent/Tools work best in scenarios where they are called upon.
  • Option 4: Hybrid Optimal (Monthly Average $5-20)recommend) For complex tasksClaude Sonnet 4.6, DeepSeek-V3.2 for everyday conversations, and heartbeat/timer functions using...Gemini For Flash or native Ollam, the fallback chain is set to Sonnet → Haiku → DeepSeek-V3.2. This is the best choice for most users, balancing performance and cost, and includes a fallback mechanism.automaticSpeed limits are handled.
  • Option 5: Completefree Option A is a local Ollam+Qwen3.5-Coder:32B or Devstral-24B (requires 32GB RAM). Option B is...freeAPI combination – GLM-4.5-Flash + ERNIE Speed+Gemini Flash memory. Suitable for privacy-sensitive, purely experimental applications. Local solutions require relatively high-end hardware.

  Security Model

  • DM pairing protectionWhen an unknown user sends a private message to your OpenClaw account via any messaging channel (WhatsApp, Telegram, etc.), the system does not process the message but returns a pairing code. The message is only processed after you manually approve it. This prevents strangers from abusing your account.Agent(And your API limit).
  • Group Sandbox ModeIn a group environment, OpenClaw runs in sandbox mode by default. Sessions within each group are isolated from each other; MEMORY.md (long-term memory) is only loaded in the main session of private chats and is not visible in groups. RequireMention can be configured to only respond when mentioned by @.
  • Tool Access ControlYou can configure either an allowlist (whitelist mode, allowing only listed tools to be used) or a denylist (blacklist mode, prohibiting listed tools). You can also completely disable the browser.automatic(Canvas visualization), (Canvas visualization), (nodes) (local device node control such as camera, screen recording).

  Known security incidents

  • CVE-2026-25253 is a remote code execution vulnerability.By forging the Origin header to bypass WebSocket authentication, attackers can connect to unauthenticated instances and execute arbitrary code.
  • ClawHavoc supply chain attackApproximately 20% of the skills on the ClawHub platform were infected with malware, affecting more than 135,000 devices.
  • Anthropic blocks OAuthAnthropic has banned [something] due to misuse issues.ClaudeSubscription accounts connect to OpenClaw via OAuth, retaining only the API Key access method.
  • Google account suspension incidentA large number of users had their entire Google accounts banned without warning due to frequent API calls by Gmail Skills that triggered abuse detection.
  • 30,000+ uncertified exposed instancesSecurity scans revealed that a large number of public OpenClaw instances had their ports completely open, posing a serious risk of data leakage and remote control.
  • Ministry of Industry and Information Technology Security Warning (March 8-9, 2026)For the first time, domestic official institutions have...open sourceAI AgentThe project published a paper pointing out that its "blurred trust boundaries" make it a high-value target for attacks.
  • Malicious npm package spoofing incident (March 2026)A fake official installer distributes the GhostLoader remote control trojan on npm, stealing user credentials and encrypted wallets.
  • ClawJacked zero-click vulnerability (early March 2026)Malicious websites can launch silent attacks through the browser, completely taking over the user's local OpenClaw instance without any user interaction.

  Cost control

  OpenClaw's out-of-control API fees mainly stem from itsOperating mechanism:Agentdeal withSimpleTasks also trigger multiple rounds of reasoning, carry a lot of context (Skills descriptions + memory system), and run 24/7, resulting in token consumption being dozens of times that of ordinary chat. A typical example is a user receiving a $1,100 bill overnight due to email processing loops.

  The core of cost control isThree-level Fallback chain:fromClaude Sonnet was downgraded to a combination of "Sonnet→Haiku→DeepSeek", allowingSimpleTaskautomaticBy adopting a low-cost model, API costs can be reduced by 80-95%.

  mustSet daily budget limit(maxTokensPerDay and maxCostPerDay), even those with ample funds can prevent [the spread of the virus].AgentThe wallet is emptied during loop reasoning; at the same time, it can be paired with local models (such as Ollam running Qwen3) to handle heartbeats and scheduled tasks. The server cost itself is very low (6-12 yuan per month). What really burns money is the uncontrolled API calls.

  Shrimp farming culture

  In 2026, with the emergence of the OpenClaw protocol,AIThe community has given rise to a unique subculture called "shrimp farming"—where users operate and maintain shrimp farms.AI AgentExamples are allowed to interact, post, and even trade stocks on dedicated social platforms such as Moltbook or the Chinese version of InStreet. These platforms are for personal use only.AI AgentThe interaction was limited to observation by humans, creating a "cyber nurturing" experience of watching digital life evolve autonomously. Its popularity peaked after Meta acquired Moltbook and launched InStreet.

  Alternative products

  • zeroclaw(24.5K Stars, written in Rust) is a lightweight autonomous...AIThe assistant infrastructure is fast to start and has low memory usage, making it suitable for resource-constrained environments.
  • nanoclaw(20.3K Stars, TypeScript) is a lightweight containerization alternative, implementing the core functionality of OpenClaw in just 4,000 lines of code, with an extremely low learning curve.
  • EasyClawFocusing on ease of use in the "last mile," it lowers the deployment threshold and is suitable for non-technical users.
  • 1Panel(34.1K Stars, Go) is the server panel.One-clickDeploy OpenClaw and manage other services on the server.
  • Umbrel(10.7K Stars, TypeScript) is a home server operating system for personal NAS/servers.One-clickInstall OpenClaw.

  Domestic ecology

  Domestic Claw Product Selection Guide

  • MaxClawMiniMax is a cloud-based product based on OpenClaw, with the default model MiniMax M2.5. Prices start from ¥39/month. Its core selling points are 18-second deployment, the lowest price, and 5-minute Lark integration.
  • AutoClaw(Zhipu)AIThis is a client-side product based on OpenClaw, with the default model GLM-5/Pony-Alpha-2.freeThe points-based system's core selling points are 96 pre-installed skills and the AutoGLM browser.automaticchange,One-clickInstall.
  • QClaw(Tencent) is a client-side product based on OpenClaw, with the default model Kimi-2.5 (which can be changed by default), currently in internal testing.freeThe core selling point is direct connection to WeChat/QQ.One-clickInstallation and data retention are local.
  • ArkClaw(Byte/Volcano) is a cloud product based on OpenClaw, with multiple models including the default Seed 2.0 model. It comes bundled with Coding Plan Pro and its core selling points are out-of-the-box usability and deep adaptation to Lark.
  • WorkBuddy(Tencent) is a client-side product, compatible with the OpenClaw Skills ecosystem, and its models support mixed-source/DeepSeek/GLM, etc.freeIts core selling point is the integration of WeChat Work/QQ.
  • LobsterAI(NetEase Youdao) isopen sourceClient-side product, not based on OpenClaw (self-developed), multiple models are available.freeopen sourceIts core selling points are its GUI interface, sandbox isolation, and strong Office capabilities.
  • CoPaw(Ali Tongyi) isopen sourceThis is a client-side + cloud-based product, not based on OpenClaw (self-developed), and supports models such as Qwen series/Ollama.freeopen sourceIts core selling points are dual-terminal and cloud deployment and multiple channels including DingTalk, Lark, and QQ.
  • miclaw(Xiaomi) is a mobile product, not based on OpenClaw (self-developed), the model is MiMo, and it is in the packaging and testing phase.freeIts core selling points are native mobile phone operation and interconnection of 1 billion+ devices in the Mi Home IoT ecosystem.

  freeGet theOpenClaw Orange BookThe original PDF report is available by scanning the QR code and replying with the following code: OpenClaw Orange Book