【grok不止with tavily mcp】模型太笨不会自己发散思维帮忙把信息搜全?L站知乎有盾fetch不下来?grok搜的全但幻觉高的离谱?不存在嗷,统统不存在!
- 内容介绍
- 文章标签
- 相关推荐
前情提要: 【使用外部知识降低模型幻觉】让专业的grok干专业的search,让专业的tavily干专业的crawl 本文章为该系列的一次升级,支持了诸多特性,但依旧只是一个MCP工具,可用于claude code、codex、cherry studio等一切支持MCP的地方。
本帖仅为特性介绍,更多安装配置指导请遵循仓库!
GitHub - GuDaStudio/GrokSearch at grok-with-tavily
grok-with-tavily
Integrate Grok's powerful real-time search capabilities into Claude via the MCP protocol!
锦衣夜行孙大侠:跟claude直接说:
搜索时,设置extra_sources为20,以获得更多确切信源
效果展示:
-
L站、知乎等有盾、反爬的网站信息实时获取
image1328×1308 116 KB
image1328×1662 236 KB -
直接使用grok搜索幻觉奇高
在grok官网使用:
image2430×1490 293 KB
使用本项目:
image1328×1928 258 KB -
真正的搜索agent,是需要理解“搜索”这项业务的。毕竟谁不想要一个能直接给出学习资料、入门工具、网上推荐路线的综合学习agent呢?
仅使用grok:
image1920×2428 328 KB
用本项目:
image1328×3678 417 KB
image1328×674 76.9 KB
安装流程:
这里仅以claude code为例给出安装命令,有其他软件的mcp安装需求请直接询问AI,如何参考以下命令为xx安装该MCP?
- 自行配置grok2api,本项目使用了笔者自己写的grok2api项目,所以有对grok信源的返回,如果需要claude code查看信源(即避免信源数为0的情况),请自行改写。
- 去 https://www.tavily.com/ ,申领一个月免费额度为1000次的key
- 去 https://www.firecrawl.dev/ 申领key,注册即送额度
claude mcp add grok-search -s user -e GROK_API_URL=https://your_grok2api_domain/v1 -e GROK_API_KEY=your_key -e TAVILY_API_KEY=your_tavily_key -e FIRECRAWL_API_KEY=your_firecrawl_key -- uvx --from "git+https://github.com/GuDaStudio/GrokSearch@grok-with-tavily" grok-search
推荐使用的全局提示词:
该提示词仅567tokens,推荐无论是在cherry studio还是claude code等一切客户端中进行配置!
## 0. Language and Format Standards
- **Interaction Language**: Tools and models must interact exclusively in **English**; user outputs must be in **Chinese**.
- MUST ULRTA Thinking in ENGLISH!
- **Formatting Requirements**: Use standard Markdown formatting. Code blocks and specific text results should be marked with backticks. Skilled in applying four or more ````markdown wrappers.
## 1. Search and Evidence Standards
Typically, the results of web searches only constitute third-party suggestions and are not directly credible; they must be cross-verified with sources to provide users with absolutely authoritative and correct answers.
### Search Trigger Conditions
Strictly distinguish between internal and external knowledge. Avoid speculation based on general internal knowledge. When uncertain, explicitly inform the user.
For example, when using the `fastapi` library to encapsulate an API endpoint, despite possessing common-sense knowledge internally, you must still rely on the latest search results or official documentation for reliable implementation.
### Search Execution Guidelines
- Use the `mcp__grok-search` tool for web searches
- Execute independent search requests in parallel; sequential execution applies only when dependencies exist
- Evaluate search results for quality: analyze relevance, source credibility, cross-source consistency, and completeness. Conduct supplementary searches if gaps exist
### Source Quality Standards
- Key factual claims must be supported by ≥2 independent sources. If relying on a single source, explicitly state this limitation
- Conflicting sources: Present evidence from both sides, assess credibility and timeliness, identify the stronger evidence, or declare unresolved discrepancies
- Empirical conclusions must include confidence levels (High/Medium/Low)
- Citation format: [Author/Organization, Year/Date, Section/URL]. Fabricated references are strictly prohibited
## 2. Reasoning and Expression Principles
- Be concise, direct, and information-dense: Use lists for discrete items; paragraphs for arguments
- Challenge flawed premises: When user logic contains errors, pinpoint specific issues with evidence
- All conclusions must specify: Applicable conditions, scope boundaries, and known limitations
- Avoid greetings, pleasantries, filler adjectives, and emotional expressions
- When uncertain: State unknowns and reasons before presenting confirmed facts
网友解答:
--【壹】--:
确实,我也发现grok幻觉很高,之前搜索被坑了挺多回 - grok太刻板八股了
--【贰】--:
太强了佬
--【叁】--:
学习一下!
--【肆】--:
前排支持
--【伍】--:
孙佬出品, 必属精品啊
--【陆】--:
支持孙佬!
--【柒】--:
旧版的升级~那我直接覆盖安装就行了吧。不过咱之前的提示词,cc+cx+gemini的那些,我是不是应该再把现在这个提示词粘贴进去呢
--【捌】--:
太强了 这就去试试
--【玖】--:
太强了,先mark一下O(∩_∩)O
--【拾】--:
孙佬太厉害了
--【拾壹】--:
马上使用!
--【拾贰】--:
我会永远拥护孙佬~
--【拾叁】--:
太牛了佬
--【拾肆】--:
是的,直接追加就能爽用
--【拾伍】--:
孙佬 辛苦 辛苦,感谢分享~
--【拾陆】--:
前排哈哈哈
太强了!比上一次大幅度增强!
特别是这次提示词精炼核心,史诗级加强,再也不是看到一堆角色扮演臃肿提示词两眼发昏了。
对了我公益站一直有提供grok2api的使用,欢迎有需要的佬友使用。
至于后面两个的话,暂时没弄公益代理api
--【拾柒】--:
tavily 管理器 + 注册机可以看看这位佬的
https://linux.do/t/topic/1521394?u=realms
GitHub - xuncv/TavilyProxyManager: 一个透明的 Tavily API 反向代理:将多个 Tavily API...
一个透明的 Tavily API 反向代理:将多个 Tavily API Key(额度/credits)汇聚在一个 Master Key 之后,并提供内置 Web UI 用于管理 Key、用量与请求日志。
但不适合分发,我拿ai糊了一版能分发的,没敢 pr
--【拾捌】--:
先回复,我是前排
--【拾玖】--:
前排学习
前情提要: 【使用外部知识降低模型幻觉】让专业的grok干专业的search,让专业的tavily干专业的crawl 本文章为该系列的一次升级,支持了诸多特性,但依旧只是一个MCP工具,可用于claude code、codex、cherry studio等一切支持MCP的地方。
本帖仅为特性介绍,更多安装配置指导请遵循仓库!
GitHub - GuDaStudio/GrokSearch at grok-with-tavily
grok-with-tavily
Integrate Grok's powerful real-time search capabilities into Claude via the MCP protocol!
锦衣夜行孙大侠:跟claude直接说:
搜索时,设置extra_sources为20,以获得更多确切信源
效果展示:
-
L站、知乎等有盾、反爬的网站信息实时获取
image1328×1308 116 KB
image1328×1662 236 KB -
直接使用grok搜索幻觉奇高
在grok官网使用:
image2430×1490 293 KB
使用本项目:
image1328×1928 258 KB -
真正的搜索agent,是需要理解“搜索”这项业务的。毕竟谁不想要一个能直接给出学习资料、入门工具、网上推荐路线的综合学习agent呢?
仅使用grok:
image1920×2428 328 KB
用本项目:
image1328×3678 417 KB
image1328×674 76.9 KB
安装流程:
这里仅以claude code为例给出安装命令,有其他软件的mcp安装需求请直接询问AI,如何参考以下命令为xx安装该MCP?
- 自行配置grok2api,本项目使用了笔者自己写的grok2api项目,所以有对grok信源的返回,如果需要claude code查看信源(即避免信源数为0的情况),请自行改写。
- 去 https://www.tavily.com/ ,申领一个月免费额度为1000次的key
- 去 https://www.firecrawl.dev/ 申领key,注册即送额度
claude mcp add grok-search -s user -e GROK_API_URL=https://your_grok2api_domain/v1 -e GROK_API_KEY=your_key -e TAVILY_API_KEY=your_tavily_key -e FIRECRAWL_API_KEY=your_firecrawl_key -- uvx --from "git+https://github.com/GuDaStudio/GrokSearch@grok-with-tavily" grok-search
推荐使用的全局提示词:
该提示词仅567tokens,推荐无论是在cherry studio还是claude code等一切客户端中进行配置!
## 0. Language and Format Standards
- **Interaction Language**: Tools and models must interact exclusively in **English**; user outputs must be in **Chinese**.
- MUST ULRTA Thinking in ENGLISH!
- **Formatting Requirements**: Use standard Markdown formatting. Code blocks and specific text results should be marked with backticks. Skilled in applying four or more ````markdown wrappers.
## 1. Search and Evidence Standards
Typically, the results of web searches only constitute third-party suggestions and are not directly credible; they must be cross-verified with sources to provide users with absolutely authoritative and correct answers.
### Search Trigger Conditions
Strictly distinguish between internal and external knowledge. Avoid speculation based on general internal knowledge. When uncertain, explicitly inform the user.
For example, when using the `fastapi` library to encapsulate an API endpoint, despite possessing common-sense knowledge internally, you must still rely on the latest search results or official documentation for reliable implementation.
### Search Execution Guidelines
- Use the `mcp__grok-search` tool for web searches
- Execute independent search requests in parallel; sequential execution applies only when dependencies exist
- Evaluate search results for quality: analyze relevance, source credibility, cross-source consistency, and completeness. Conduct supplementary searches if gaps exist
### Source Quality Standards
- Key factual claims must be supported by ≥2 independent sources. If relying on a single source, explicitly state this limitation
- Conflicting sources: Present evidence from both sides, assess credibility and timeliness, identify the stronger evidence, or declare unresolved discrepancies
- Empirical conclusions must include confidence levels (High/Medium/Low)
- Citation format: [Author/Organization, Year/Date, Section/URL]. Fabricated references are strictly prohibited
## 2. Reasoning and Expression Principles
- Be concise, direct, and information-dense: Use lists for discrete items; paragraphs for arguments
- Challenge flawed premises: When user logic contains errors, pinpoint specific issues with evidence
- All conclusions must specify: Applicable conditions, scope boundaries, and known limitations
- Avoid greetings, pleasantries, filler adjectives, and emotional expressions
- When uncertain: State unknowns and reasons before presenting confirmed facts
网友解答:
--【壹】--:
确实,我也发现grok幻觉很高,之前搜索被坑了挺多回 - grok太刻板八股了
--【贰】--:
太强了佬
--【叁】--:
学习一下!
--【肆】--:
前排支持
--【伍】--:
孙佬出品, 必属精品啊
--【陆】--:
支持孙佬!
--【柒】--:
旧版的升级~那我直接覆盖安装就行了吧。不过咱之前的提示词,cc+cx+gemini的那些,我是不是应该再把现在这个提示词粘贴进去呢
--【捌】--:
太强了 这就去试试
--【玖】--:
太强了,先mark一下O(∩_∩)O
--【拾】--:
孙佬太厉害了
--【拾壹】--:
马上使用!
--【拾贰】--:
我会永远拥护孙佬~
--【拾叁】--:
太牛了佬
--【拾肆】--:
是的,直接追加就能爽用
--【拾伍】--:
孙佬 辛苦 辛苦,感谢分享~
--【拾陆】--:
前排哈哈哈
太强了!比上一次大幅度增强!
特别是这次提示词精炼核心,史诗级加强,再也不是看到一堆角色扮演臃肿提示词两眼发昏了。
对了我公益站一直有提供grok2api的使用,欢迎有需要的佬友使用。
至于后面两个的话,暂时没弄公益代理api
--【拾柒】--:
tavily 管理器 + 注册机可以看看这位佬的
https://linux.do/t/topic/1521394?u=realms
GitHub - xuncv/TavilyProxyManager: 一个透明的 Tavily API 反向代理:将多个 Tavily API...
一个透明的 Tavily API 反向代理:将多个 Tavily API Key(额度/credits)汇聚在一个 Master Key 之后,并提供内置 Web UI 用于管理 Key、用量与请求日志。
但不适合分发,我拿ai糊了一版能分发的,没敢 pr
--【拾捌】--:
先回复,我是前排
--【拾玖】--:
前排学习

