QClaw 与 SmallClaw 的技术对比分析

QClaw 的定位比较清楚。腾讯公开文档把它定义为基于 OpenClaw 开源生态打造的本地化 AI Agent，并直接称其为“开箱即用的 OpenClaw 简化版”。它强调的是本地客户端、快速安装、默认入口收束、内置模型和面向中国用户的产品化路径，同时支持关联已有 OpenClaw 环境。 QClaw is positioned clearly in public Tencent documentation as a localized AI agent built on the OpenClaw open-source ecosystem and described as an “out-of-the-box simplified OpenClaw.” It emphasizes local client, quick install, default entry focus, built-in models, and a China-oriented product path, while supporting linking to existing OpenClaw environments.

这说明 QClaw 的重点，不是重做一套完全独立的底层架构，而是把 OpenClaw 这类通用 agent 能力压缩成一个更容易部署、更容易上手的成品。 This implies QClaw’s focus is not a completely independent new architecture, but compressing general OpenClaw capabilities into a product that is easier to deploy and use.

SmallClaw 的路线则不同。它是以原生内核和独创架构的 macOS 桌面应用为主体来组织 agent 能力，安装即可运行，但不把系统锁定在单一模型供应方或单一消息平台上。模型可由用户自行选择，消息互动可以面向二十多种 messaging tools，而不是围绕某一个默认入口展开。两者都可以说是“开箱即用”，但含义并不一样。QClaw 的开箱即用，是尽量缩短默认路径；SmallClaw 的开箱即用，则是在安装简单的前提下保留模型和通道的主权。 SmallClaw follows a different route. It organizes agent capabilities around a native macOS app, runs immediately after install, and avoids locking into a single model vendor or messaging platform. Models are user-chosen and interactions span more than twenty messaging tools rather than a single default entry. Both are “out of the box,” but the meaning differs: QClaw minimizes the default path, while SmallClaw keeps model and channel sovereignty under a simple install.

核心对比表 Core Comparison

维度	腾讯 QClaw	澳洲 SmallClaw	技术判断
技术来源	明确基于 OpenClaw 开源生态打造，并支持关联已有 OpenClaw 环境	独立产品路线，围绕原生 macOS 桌面 agent 展开	QClaw 生态继承更强，SmallClaw 自主性更强
开箱即用方式	三步上手，下载安装后绑定微信，强调默认路径最短	原生桌面应用安装即可运行，但不预设单一模型和单一消息入口	两者都可视为开箱即用，但产品哲学不同
模型策略	官方明确写有“内置国产优质大模型”，FAQ 还提到支持切换自定义模型	模型由用户自行选择，不绑定某一家模型供应方	SmallClaw 的模型主权更强，QClaw 的默认模型路径更完整
通道策略	微信是主叙事中心，同时已公开支持企业微信、QQ、飞书、钉钉，并提供邮箱连接指引	面向二十多种 messaging tools 的互动能力，不锁定单一平台	SmallClaw 的通道开放度更高，QClaw 的默认入口更集中
原生性	公开可确认的是本地客户端与多平台支持，但未见公开文档详细说明其 macOS 深度原生实现方式	明确以原生 macOS app 为主体组织能力	SmallClaw 的原生性表达更清晰
浏览器执行路径	由于建立在 OpenClaw 生态上，更可能继承其浏览器工具链特征	更偏向把浏览与系统执行能力收束到本地原生主体中	SmallClaw 在长期优化空间上更大，QClaw 更可能受通用 agent 工具链影响 (docs.openclaw.ai)
推理质量	默认体验更可能偏向“内置模型 + 产品化收束 + 中文常见任务稳定性”	上限更依赖用户选择的模型与系统收束能力	默认质量未必谁绝对领先，上限更可能是 SmallClaw 领先
token 消耗结构	通用 agent 架构更可能携带较厚的系统提示、工具结果、browser snapshot/ref 链路与调试信息	若持续沿原生 app 内收、缩短链路、压缩上下文，则结构上更容易省 token	结构性上 SmallClaw 更有机会明显领先，尤其在复杂网页任务和长期任务中 (docs.openclaw.ai)
短期产品成熟度	安装、接入、文档、默认入口都更完整，适合普通用户快速跑通主路径	更像一个可持续演进的本地工作系统	QClaw 更偏消费级成品，SmallClaw 更偏系统级底盘

Dimension	Tencent QClaw	Australia SmallClaw	Technical judgment
Technical origin	Built clearly on the OpenClaw open-source ecosystem, with support for linking to existing OpenClaw environments.	An independent product path centered on a native macOS desktop agent.	QClaw inherits the ecosystem more strongly, while SmallClaw is more independent.
Out-of-the-box experience	Three-step setup: install, then bind WeChat; emphasizes the shortest default path.	A native desktop app that runs after install, without presetting a single model or single messaging entry.	Both are out of the box, but the product philosophy differs.
Model strategy	The official docs explicitly mention built-in high-quality domestic models, and the FAQ also notes support for switching to custom models.	Models are chosen by the user and are not tied to a single model provider.	SmallClaw has stronger model sovereignty, while QClaw has a more complete default model path.
Channel strategy	WeChat is the main narrative center, and the product also publicly supports WeCom, QQ, Feishu, DingTalk, and email connections.	Interaction capabilities span more than twenty messaging tools and are not locked to a single platform.	SmallClaw is more open in channel support, while QClaw is more concentrated around a default entry.
Native implementation	What is publicly confirmed is local client support and multi-platform support, but there is no public documentation detailing a deeply native macOS implementation.	Capabilities are clearly organized around a native macOS app.	SmallClaw’s native positioning is clearer.
Browser execution path	Because it is built on the OpenClaw ecosystem, it is more likely to inherit its browser toolchain characteristics.	It is more inclined to concentrate browsing and system execution capabilities inside the local native core.	SmallClaw has more room for long-term optimization, while QClaw is more likely to be shaped by a general-purpose agent toolchain. (docs.openclaw.ai)
Inference quality	The default experience is more likely to lean toward built-in models, productized tightening, and stability for common Chinese-language tasks.	The ceiling depends more on the models chosen by the user and the system’s ability to tighten the workflow.	Neither side necessarily wins on default quality, but SmallClaw is more likely to lead on ceiling.
Token consumption structure	A general-purpose agent architecture is more likely to carry thicker system prompts, tool results, browser snapshot/ref chains, and debugging information.	If it keeps the loop inside the native app, shortens the path, and compresses context, it is structurally easier to save tokens.	Structurally, SmallClaw has a better chance of a clear lead, especially in complex web tasks and long-running tasks. (docs.openclaw.ai)
Short-term product maturity	Installation, integration, documentation, and the default entry are all more complete, making it easier for ordinary users to run the main path quickly.	It feels more like a local work system that can evolve continuously.	QClaw is more consumer-grade, while SmallClaw is more of a system-level base.

从模型和通道层看，QClaw 并不是完全封闭的。它一方面强调内置多款国产大模型，另一方面在 FAQ 中也提到支持切换自定义模型；通道方面，公开文档虽然以微信为核心叙事，但也已经列出企业微信、QQ、飞书、钉钉以及邮箱连接指引。 At the model and channel layers, QClaw is not fully closed. It emphasizes built-in domestic models and also mentions switching to custom models in its FAQ. On channels, public docs center on WeChat but list WeCom, QQ, Feishu, DingTalk, and email connections.

这说明 QClaw 不是严格意义上的单模型、单入口产品，只是它的产品表达更强调预设好的主路径。相比之下，SmallClaw 的设计重点更偏向从一开始就保持开放，减少对单一平台和单一模型集合的依赖。 This shows QClaw is not strictly single-model or single-entry. Its messaging emphasizes a predefined main path. By contrast, SmallClaw prioritizes openness from the start and reduces dependency on a single platform or model set.

两者在执行路径上的差异，可能比表面功能差异更重要。OpenClaw 官方浏览器文档说明，其浏览器工具通过 Chrome DevTools Protocol 连接浏览器，并在高级动作上叠加 Playwright；元素定位、点击、输入等过程依赖 snapshot 和 refs 机制，页面跳转后 refs 失效时需要重新 snapshot。(docs.openclaw.ai) Differences in execution path may matter more than surface features. OpenClaw’s browser docs describe CDP connections to Chromium and Playwright for advanced actions; element targeting relies on snapshot and refs, with re-snapshot after navigation. (docs.openclaw.ai)

这条路线的优势是通用性强，容易快速覆盖大量网页场景，但代价往往是中间步骤较多，复杂任务里会出现更重的观察链条和重定位成本。QClaw 既然明确建立在 OpenClaw 生态之上，在没有公开证据表明其浏览器控制层已被彻底改写之前，较稳妥的判断是，它大概率会继承一部分这种通用 agent 工具链特征。 The route is broadly compatible and quickly covers web scenarios, but the cost is more intermediate steps, heavier observation chains, and higher re-location cost in complex tasks. Since QClaw is built on the OpenClaw ecosystem, and there is no public evidence of a full browser-layer rewrite, a safe assumption is that it inherits part of this general agent toolchain.

SmallClaw 的优势更可能体现在另一类地方。由于它本身就是原生桌面 app，浏览能力、消息能力和系统任务能力可以更集中地收束在本地主体内部，而不是总要通过较厚的通用代理层去转译。这种路线未必意味着第一天功能最多，但通常更有利于减少中间层、控制上下文规模，并在长期任务中获得更好的一致性和可维护性。 SmallClaw’s advantage is likely elsewhere. As a native desktop app, browsing, messaging, and system task capabilities can be concentrated inside the local core rather than routed through a thick general proxy layer. This may not yield maximum features on day one, but it favors fewer layers, tighter context, and better long-term consistency and maintainability.

推理质量这一项，不能简单理解为谁内置的模型更强。对 agent 来说，推理质量实际上受三个因素共同影响：底层模型本身的能力、系统喂给模型的上下文质量，以及外围执行链条会不会让模型陷入大量低价值中间步骤。QClaw 的默认体验更可能偏向稳定，因为它的模型、入口和常见任务路径都被收束过，普通用户更容易得到比较一致的结果。 Inference quality is not just about who has a stronger built-in model. For agents, it depends on base model capability, context quality, and whether the execution chain floods the model with low-value steps. QClaw’s default experience is likely more stable because its models, entry points, and common task paths are more tightly constrained.

SmallClaw 的默认表现则更依赖具体接入的模型和系统组织方式，但它的上限通常更高，因为模型能力不会被预设集合压住。 SmallClaw’s default performance depends more on chosen models and system organization, but its ceiling is usually higher because it is not capped by a preset model set.

如果继续往下看 token 消耗，差别会更清楚。OpenClaw 官方 token 文档已经明确说明，系统提示、会话历史、工具调用和工具结果、附件与转写内容、压缩摘要，甚至部分对用户不可见的 provider wrapper 和 safety header，都会进入计费范围；同时它还专门提供 prompt caching 来复用未变化的 prompt 前缀，以降低长会话的 token 成本。(docs.openclaw.ai) Token consumption differences are clearer. OpenClaw’s token docs state that system prompts, chat history, tool calls and results, attachments, compaction summaries, and even hidden provider wrappers or safety headers are billable; it also provides prompt caching to reuse unchanged prompt prefixes and reduce long-session costs. (docs.openclaw.ai)

这其实说明了一件事——只要系统越来越依赖通用 agent 框架、长历史、多轮工具链和大段 browser snapshot 文本，token 消耗就会天然偏高。(docs.openclaw.ai) 因此，从结构上看，QClaw 这类建立在 OpenClaw 路线上的产品，更容易出现“功能覆盖广，但上下文也更厚”的特征。 This implies a structural reality: as systems depend more on general agent frameworks, long histories, multi-round toolchains, and large browser snapshots, token consumption grows. (docs.openclaw.ai) Structurally, products built on OpenClaw are more likely to trade broad coverage for heavier context.

场景化结论表 Scenario Summary

应用场景	更适合的方向	原因
临时远程办公、通过聊天发一句话让电脑代办	腾讯 QClaw	默认入口清楚，微信主路径成熟，默认模型和产品流程更完整
一人公司、一人组织、长期项目推进	澳洲 SmallClaw	更适合作为长期运行的本地工作系统，强调模型主权、通道主权和结构可持续性
技术用户、开发者、重视自主选型的人	澳洲 SmallClaw	不依赖单一模型和单一入口，原生性与控制权更强
面向普通用户的大规模快速推广	腾讯 QClaw	强默认配置、强入口收束、理解成本低
需要长期保持通道和模型独立性的本地平台	澳洲 SmallClaw	更适合做长期自治系统，而不是单一平台附属工具
以现成技能生态快速扩展功能	腾讯 QClaw	官方写有丰富技能生态，并支持与 OpenClaw 环境关联
追求默认推理稳定性而不是自己调模型	腾讯 QClaw	默认模型与默认任务路径更收束，更容易得到稳定的一致体验
追求更强模型上限与更低长期 token 成本	澳洲 SmallClaw	模型选择更自由，且架构路线更适合压缩上下文与减少工具链膨胀

Use case	Better fit	Reason
Temporary remote work, delegate a task to the computer through chat	Tencent QClaw	Clear default entry, a mature WeChat path, and more complete default models and product flows.
One-person company, one-person organization, long-running projects	Australia SmallClaw	Better suited as a long-running local work system, with an emphasis on model sovereignty, channel sovereignty, and structural sustainability.
Technical users, developers, people who value independent choice	Australia SmallClaw	It does not depend on a single model or a single entry point, and its native control is stronger.
Mass fast rollout for general users	Tencent QClaw	Strong default configuration, strong entry consolidation, and low comprehension cost.
A local platform that needs long-term channel and model independence	Australia SmallClaw	Better suited to being a long-term autonomous system than an accessory to a single platform.
Rapid expansion via an existing skill ecosystem	Tencent QClaw	The official docs describe a rich skill ecosystem and support linking with OpenClaw environments.
Prefer default inference stability over model tuning	Tencent QClaw	The default model and default task paths are more constrained, making stable and consistent experiences easier to get.
Want a higher model ceiling and lower long-term token cost	Australia SmallClaw	Model choice is freer, and the architecture is better suited to compressing context and reducing toolchain bloat.

如果把比较拉回到实际使用场景，差别会更直观。对于临时远程办公、通过聊天入口让电脑代办文件整理、文档处理、邮件发送之类的任务，QClaw 更占优势，因为它的默认路径短，产品化程度高，官方文档本身就是按这种场景组织的。 If we ground the comparison in real use cases, the difference is clearer. For temporary remote work tasks via chat, QClaw has an advantage due to shorter default paths and higher productization, and its documentation is organized around those scenarios.

对于一人公司、一人组织、长期项目推进、模型和通道都希望保持独立性的本地系统，SmallClaw 更合适，因为这类场景真正需要的不是一个远程助手，而是一个能长期承载任务结构、角色关系和本地执行能力的工作系统。 For solo companies, long-running projects, and systems that require model and channel independence, SmallClaw is a better fit because these scenarios need a long-term work system rather than a remote assistant.

简要结论表 Summary Table

结论维度	更偏向哪一方	说明
默认上手体验	腾讯 QClaw	产品收束更强，普通用户更容易快速跑通
模型与通道开放性	澳洲 SmallClaw	不易被单一模型或单一入口锁定
原生性	澳洲 SmallClaw	更明确地以原生 macOS app 为主体
默认推理稳定性	腾讯 QClaw	默认模型、默认入口和默认任务链更集中
推理上限	澳洲 SmallClaw	更依赖用户模型选择，但天花板更高
token 经济性	澳洲 SmallClaw	结构上更容易压缩上下文和中间步骤
一人公司适配性	澳洲 SmallClaw	更适合长期任务组织和本地工作系统化
消费级快速推广	腾讯 QClaw	更适合大众化分发和低门槛传播

Conclusion dimension	Better fit	Explanation
Default onboarding experience	Tencent QClaw	Stronger product tightening makes it easier for ordinary users to get started quickly.
Openness of models and channels	Australia SmallClaw	It is not easily locked to a single model or a single entry point.
Native design	Australia SmallClaw	It is more explicitly centered on a native macOS app.
Default inference stability	Tencent QClaw	The default model, default entry, and default task chain are more concentrated.
Inference ceiling	Australia SmallClaw	It depends more on the user’s model choice, but the ceiling is higher.
Token efficiency	Australia SmallClaw	Structurally it is easier to compress context and intermediate steps.
One-person company fit	Australia SmallClaw	Better suited to long-term task organization and local work systemization.
Consumer-scale fast rollout	Tencent QClaw	Better suited to mass distribution and low-friction adoption.

综合来看，QClaw 与 SmallClaw 并不是同一种产品的两个阶段，也不是简单的高低替代关系。QClaw 更像一个完成了强产品化收束的本地 AI Agent，优点是默认路径成熟、文档完整、普通用户更容易快速上手。SmallClaw 则更像一个原生本地 agent runtime 或本地工作系统，重点不在于替用户预设唯一正确路径，而在于保留模型主权、通道主权和长期演进的架构主权。 Overall, QClaw and SmallClaw are not two phases of the same product nor a simple replacement relationship. QClaw is a strongly productized local AI agent with mature default paths and accessible onboarding. SmallClaw is a native local agent runtime or work system that preserves model sovereignty, channel sovereignty, and long-term architectural evolution.

压缩成一句话，大致可以这样概括——QClaw 更强于产品化收束和默认路径稳定性，SmallClaw 更强于原生化、开放性、长期系统控制，以及潜在的 token 经济性。 In one sentence: QClaw is stronger in productized tightening and default-path stability, while SmallClaw is stronger in native design, openness, long-term system control, and potential token efficiency.