Skip to main content
Glama
kimi-search-mcp
by Wildare-98

kimi-search-mcp

基于 Node.js 的本地 stdio MCP Server,将 Kimi 开放平台的联网搜索 Basic、联网搜索 Pro 和网页抓取 REST API 封装为三个 MCP 工具。

工具

工具

Kimi API

说明

kimi_search

POST /v1/tools/search

返回标题、URL、摘要、日期和来源权威性,可选网页正文

kimi_search_pro

POST /v1/tools/search_pro

返回相关正文片段,支持站点与发布时间过滤,适合问答、RAG 和研究

kimi_fetch

POST /v1/tools/fetch

抓取指定 HTTP/HTTPS 网页,返回标题和 Markdown 正文

工程特性:

  • 纯本地 stdio 模式,无需部署 HTTP 服务

  • API Key 只从环境变量 MOONSHOT_API_KEY 读取

  • 408、429 和部分 5xx 错误自动重试,最多 3 次

  • 搜索超时参数为 1–60 秒,客户端额外预留 10 秒网络时间

  • 日志只写入 stderr,不干扰 stdio MCP 协议

Related MCP server: web-tools-mcp

要求

安装

git clone https://github.com/Wildare-98/kimi-search-mcp.git
cd kimi-search-mcp
npm install

Kimi Code 配置

在用户级 ~/.kimi-code/mcp.json 或项目级 .kimi-code/mcp.json 中加入:

{
  "mcpServers": {
    "kimi-search": {
      "command": "node",
      "args": ["D:/AIFile/kimicode/Chatbot/kimi-search-mcp/index.js"],
      "env": {
        "MOONSHOT_API_KEY": "sk-你的key"
      }
    }
  }
}

将 args 中的路径替换为实际绝对路径。也可以把 MOONSHOT_API_KEY 设为系统环境变量,然后删除配置中的 env 段,避免把密钥写入配置文件。

新增配置只会注册到新会话。保存后新建会话,使用 /mcp 检查连接状态。

其他支持相同 JSON 配置的 MCP 客户端也可使用该配置,只需调整绝对路径。

使用建议

  1. 普通搜索先调用 kimi_search,保持 include_content=false。

  2. 需要直接获取相关段落、限定站点或时间时,调用 kimi_search_pro。

  3. 已知 URL 或需要阅读完整网页时,调用 kimi_fetch。

搜索查询应尽量包含明确的实体、时间和来源限定词。kimi_search_pro 的 time_window 日期格式支持 YYYY、YYYY-MM 和 YYYY-MM-DD,sites 最多包含 5 个站点。

可选环境变量

变量

默认值

说明

MOONSHOT_API_KEY

无,必填

Kimi API Key

MOONSHOT_BASE_URL

https://api.moonshot.cn/v1

API 基础地址,用于兼容代理或测试

计费

三个接口按成功且非空的结果分别计费;失败、超时或空结果不计费。详见联网搜索定价。

API 文档

Available Tools

3 tools
kimi_fetchKimi Web FetchA

Fetch a single HTTP/HTTPS web page with Kimi and return its title and Markdown content. Use after kimi_search to read a specific URL. Subject to site availability, access controls, and Kimi security policy.

Inputs: url (required).

kimi_search_proKimi Web Search ProA

Search the web with Kimi Search Pro. Returns relevant page excerpts ranked by relevance, with source URLs and authority. Best for web QA, RAG, research, and reports. Supports site and publication-date filters.

Inputs: text_query (required), limit, sites, time_window, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.9/5.0

    Scored across 3 tools

    Disambiguation4/5

    kimi_fetch is clearly distinct from the two search tools. kimi_search and kimi_search_pro overlap heavily as web search operations, though the descriptions differentiate basic overview results from relevance-ranked excerpt results with filters.

    Naming Consistency5/5

    All tools use a consistent kimi_ prefix with snake_case action names: kimi_search, kimi_search_pro, and kimi_fetch. The naming is predictable and easy to map to the intended operation.

    Tool Count5/5

    Three tools are well-scoped for a web search and page-fetching server: two search modes plus a fetch operation. No unnecessary tools are present for the apparent purpose.

    Completeness4/5

    The server covers web search, advanced search, and single-page fetching, which supports the core search-and-read workflow. A batch fetch or multi-URL operation is missing, but agents can work around this by calling kimi_fetch per URL.

    Maintenance

    ActivityMaintained
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      F
      maintenance
      Enables agents to perform web searches and fetch page content through Kimi Code's search/fetch capabilities via standard MCP tools.
      1
      -
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables web searching and fetching pages converted to Markdown through local stdio MCP tools, so Claude Code and other MCP clients connected via third-party relays retain internet access instead of failing with a server-side tool protocol error. Includes paginated reading of long articles, SSRF/private-network blocking, and charset handling for legacy Chinese sites.
      85 npm
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      Enables MCP-capable AI clients to run web searches through DeepSeek's search engine and fetch individual webpages as Markdown. Search results include relevance scores, authority, and publication times, and support advanced query syntax such as site:, lang:, and date filters.
      2
      51 npm
      Apache 2.0