Loading ToolBench
A professional desktop application for managing and executing AI agents with real-time chat, session management, and skills browsing. Includes CLI, Gateway server, and Electron desktop UI.
OpenClawX has 11 tools with mixed quality. Most tools have descriptions in Chinese and reasonable parameter schemas, but there are significant gaps in English documentation clarity, error handling guidance, and output schema documentation. Parameter descriptions exist but are often vague or language-dependent. No visible error recovery patterns or confirmation workflows for destructive operations. Tools like `browser` have reasonable complexity but lack structured output documentation. The server appears functional but falls below production-grade standards due to missing error handling patterns, undocumented outputs, and reliance on Chinese descriptions that may not optimize for LLM reasoning in multilingual contexts.
在系统中新增一个收藏标签。当用户要保存链接到某个尚不存在的标签（如「美女」「技术」「待读」）时，先调用本工具创建该标签，再调用 get_bookmark_tags 获取最新列表并用 save_bookmark 保存。若标签已存在会提示无需重复创建。
Control a headless browser for web automation. Actions: navigate (go to URL), snapshot (get page content with refs), screenshot (capture image), click (click element), type (type text), fill (clear and fill input), scroll (scroll page), extract (get text from element), wait (wait for element), download (download file from URL or by clicking), back/forward (navigation), close (close browser). Use refs from snapshot (e.g., @e1, @e2) to interact with elements reliably.
根据用户意图创建一个新的智能体配置。用户可在对话中通过 list_agents 与 switch_agent 选择使用。需提供智能体名称；工作空间名与系统提示词可选，未提供时自动生成。
获取系统中已维护的收藏标签列表。在用户要求保存链接时，应先调用本工具获取可用标签，再根据用户意图匹配最合适的标签名，并用 save_bookmark 保存。
将指定地址的技能安装到当前智能体工作区目录。在 OpenBot 中为用户安装技能时，应使用本工具而非 npx skills add，以保证安装到当前智能体对应的技能目录。
列出当前可用的所有智能体（id 与名称）。用户可根据列表选择并让助手调用 switch_agent 切换到对应智能体。
No documented output schemas for any tool. LLMs cannot plan downstream operations or extract specific fields if return types are not documented.
No error handling guidance or recovery patterns. Descriptions lack 'what to do if this fails' context. E.g., add_bookmark_tag says 'will prompt if tag exists' but no guidance on handling that error condition.
Destructive/write operations (add_bookmark_tag, save_bookmark, create_agent, install_skill, switch_agent, save_experience) lack dry-run or confirmation patterns. No tool annotations (destructiveHint) visible.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 55 | 2026-07-28+ | v2 |
| 2026-03-09 | C | 63 | - | v1 |
从长期记忆中语义检索与查询相关的经验总结和对话摘要。在回答与过往工作、决定、日期、人、偏好、待办、复杂任务、定时任务或需要历史经验相关的问题前，应先调用本工具获取相关记忆再作答。
将用户提供的 URL 保存到收藏库，并可关联一个或多个标签。标签名须为系统已有标签（get_bookmark_tags 返回）；若缺少某标签可先用 add_bookmark_tag 新增。
在对话结束时调用，将本轮对话中总结出的经验、要点或可复用的结论存入长期记忆，便于后续会话检索。
将当前对话切换为使用指定智能体。切换后，用户下一次及后续消息将由该智能体处理。可用于在对话中更换助手（如从主智能体切换到专项智能体）。
联网搜索：根据查询词从互联网获取最新结果（标题、链接、摘要）。无 API Key 时使用 DuckDuckGo；配置 Brave API Key 后可选用 Brave。在回答需要实时信息、新闻、天气、概念解释等问题前可调用本工具。
Descriptions are in Chinese, which may reduce LLM comprehension in English-first models. Even for multilingual models, descriptions should clarify context ('When should you call this vs. similar tools?') which is present but minimal.
browser tool has 12+ action types in enum, but no guidance on which action sequences are safe, which may fail, or what refs look like. Output schema for snapshot action (with refs) not documented.
Parameter descriptions lack validation rules. E.g., 'tagName' in add_bookmark_tag has no length, format, or character restrictions stated. 'count' in web_search accepts 1-10 but LLM doesn't know if <1 or >10 will error.
get_bookmark_tags, list_agents, web_search return lists but no pagination or result limit documented. get_bookmark_tags and list_agents may scale poorly; web_search caps at 10 but no guidance on exceeding limit.