Agentsearch Mcp

io.github.IsaiahDupree由 GitHub 帳號 IsaiahDupree 發布v1.0.0更新於 Oct 24, 2026

LLM-ready web search + instant answers + URL-to-clean-text fetch for agents and RAG.

已驗證Streamable HTTP可網頁執行網頁搜尋與抓取知識與記憶

概覽

AI 產生的概覽

遠端 MCP 連接器,為代理與 RAG 提供網頁搜尋、免金鑰即時問答,以及把網址轉成乾淨文字的擷取能力。

可以做什麼
  • 透過可切換供應商的後端搜尋網頁,回傳含位置、標題、網址、摘要與網域的標準化結果。
  • 以免金鑰的即時問答查詢定義、實體與快速事實。
  • 擷取任意公開網址,回傳去除樣板內容的純文字或 Markdown,方便放入大型語言模型上下文。
適用情境
適合需要即時網頁結果、快速事實查詢,或需要可讀頁面正文用於檢索與摘要的情境,不必自行建置擷取流程。它只是對計量上游 API 的輕量轉送,適合輕度探索使用,不適合高流量正式環境。
執行需求
需要一個遠端 streamable-HTTP 端點;MCP 呼叫端不需帳號、OAuth 或 API 金鑰,上游代理金鑰由連接器在伺服器端持有。擷取網頁需要對外網路存取。
安裝前請注意
擷取的頁面內容與搜尋結果來自第三方並原樣回傳,應視為不可信輸入。依 IP 有記憶體內的軟性速率限制,營運方稱這是付費 API 之上的免費體驗層,大量使用應走付費管道。所有工具皆為唯讀。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Agentsearch Mcp,將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable,透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式,或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南,先向你展示發布者、端點和登入方式,經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/io.github.IsaiahDupree%2Fagentsearch-mcp/install.md,並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前,先告訴我它的發布者、端點和登入方式,等我確認。不要把 API 金鑰寫進命令或我會提交的檔案;問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "agentsearch": {
      "type": "http",
      "url": "https://agentsearch-mcp.vercel.app/mcp"
    }
  }
}

README

agentsearch-mcp

A remote MCP (Model Context Protocol) connector for AgentSearch API — an LLM/MCP-native web toolkit for agents and RAG pipelines: web search (provider-abstracted SERP), keyless instant answers, and URL → clean text/markdown fetch ready for an LLM context window.

Live: https://agentsearch-mcp.vercel.app/mcp — 3 tools, one per AgentSearch /v1 endpoint. Since the upstream AgentSearch deployment is metered (RapidAPI/Apify) and gates /v1/* behind a RapidAPI proxy-secret guard, this connector authenticates its own outbound calls with that same secret (AGENTSEARCH_MCP_PROXY_SECRET, sent as the X-RapidAPI-Proxy-Secret header) and applies a soft per-IP rate limit (AGENTSEARCH_MCP_RATE_LIMIT, default 30 tool-calls/hour, in-memory) so the free MCP tier stays a discovery channel rather than an unmetered bypass of the paid listing — see lib/ratelimit.js.

What this is, and why it's a separate connector

AgentSearch API is a plain REST API. Any HTTP client can already call it directly. This repo exists because MCP clients (Claude, ChatGPT, and other MCP-aware agents) don't consume arbitrary REST APIs — they consume MCP tools. agentsearch-mcp is a thin adapter layer that:

  • Exposes each AgentSearch endpoint as a discoverable, typed MCP tool (name, description, zod input schema, annotations) that an LLM can reason about and call directly, instead of having to be taught the REST surface out-of-band.
  • Speaks the MCP streamable-HTTP transport at a single /mcp endpoint, so it can be registered as a connector in Claude, ChatGPT, or any other MCP client with one URL.
  • Does nothing else. It has no business logic of its own — every tool call is a pass-through fetch to AgentSearch, and the JSON response AgentSearch returns is handed back verbatim as the tool result.

Free discovery tier over a metered API

This connector is a free discovery/growth tier in front of the metered AgentSearch API. The upstream (agentsearch-api.vercel.app) is sold on RapidAPI/Apify; this MCP wrapper authenticates to it with the shared RapidAPI proxy secret and caps usage per-IP so it stays a taste-test rather than an unmetered path around the paid plans. Heavy/production volume should go through AgentSearch on RapidAPI/Apify.

Authentication: None on the MCP side (deliberate)

AgentSearch's data has no per-user dimension — it's public web data (SERP results, DuckDuckGo instant answers, cleaned page text). There is nothing to gate per-caller, so this connector intentionally ships with:

  • No OAuth, no login, no bearer tokens for the MCP caller
  • No Supabase / database
  • No demo-vs-real account split — every caller gets the same real, live data

api/mcp.js builds a fresh, stateless McpServer per request and serves it with zero MCP-caller auth checks. The only outbound auth is the upstream RapidAPI proxy secret described above, which the connector holds server-side.

Tool list

One tool per AgentSearch /v1 endpoint (from agentsearch-api/openapi.yaml; /api/health is intentionally not wrapped):

ToolAgentSearch endpointDescription
web_searchGET /v1/searchWeb SERP via a provider-abstracted backend (Brave or Serper). Returns normalized results with position/title/url/snippet/domain.
instant_answerGET /v1/answerKeyless DuckDuckGo instant answer — definitions, entities, quick facts.
fetch_urlGET /v1/fetchKeyless, SSRF-guarded URL → clean, boilerplate-free text or markdown for RAG.

All 3 tools are read-only and annotated { readOnlyHint: true, destructiveHint: false, idempotentHint: true, openWorldHint: true } — none of them write anything, and all of them reflect live, externally-changing web data.

How it wraps agentsearch-api

Each tool handler does a plain fetch(\${AGENTSEARCH_MCP_API_BASE_URL}${path}`, ...)` against the real AgentSearch REST API and returns the parsed JSON as MCP tool-result content:

js
{ content: [{ type: 'text', text: JSON.stringify(result, null, 2) }] }

The outbound request carries X-RapidAPI-Proxy-Secret: <AGENTSEARCH_MCP_PROXY_SECRET> so it passes the upstream guard on /v1/*. Upstream HTTP errors (4xx/5xx) are caught and surfaced as a typed MCP error result via an asError helper rather than crashing the request.

Environment variables

VarDefaultPurpose
AGENTSEARCH_MCP_API_BASE_URLhttps://agentsearch-api.vercel.appUpstream AgentSearch base URL (override for local/self-hosted testing).
AGENTSEARCH_MCP_PROXY_SECRET(empty)RapidAPI proxy secret, sent as X-RapidAPI-Proxy-Secret to pass the upstream /v1/* guard. Without it, guarded routes return 403.
AGENTSEARCH_MCP_RATE_LIMIT30Soft per-IP tools/call cap per hour (in-memory, per serverless instance).

Project layout

api/mcp.js       MCP endpoint (StreamableHTTPServerTransport, stateless, no MCP-caller auth)api/health.js    GET /api/healthlib/tools.js     All 3 tool definitions (zod schemas + fetch-and-forward handlers, proxy-secret auth)lib/ratelimit.js Soft per-IP tools/call rate limiterlocal-server.js  Plain-Node http server for local dev / smoke testing (not deployed)test/smoke.mjs   Real end-to-end smoke test (initialize, tools/list, tools/call)vercel.json      Routes /mcp -> api/mcp.js, /health -> api/health.jsserver.json      MCP registry manifest

Local development

bash
npm installnpm run dev          # starts local-server.js on :3900npm run smoke        # runs test/smoke.mjs against the local server

To exercise a real end-to-end fetch through the upstream guard, set the proxy secret first:

bash
export AGENTSEARCH_MCP_PROXY_SECRET=<the RapidAPI proxy secret>npm run dev &npm run smoke

來源:README.md,提交 c585fcf

工具

3

instant_answer

Instant answer (keyless)

DuckDuckGo Instant Answer API — definitions, entities, and quick facts. Returns a normalized answer with heading, type (abstract/answer/disambiguation), source, sourceUrl, optional image/definition, and relatedTopics, plus a meta envelope. Not a full web SERP; use web_search for that. Works with no API key.

輸入參數q必填

fetch_url

Fetch a URL as clean text/markdown (RAG-ready)

Fetches any public http(s) URL, strips boilerplate (scripts, nav, footer, ads), and returns clean text or markdown ready for an LLM context window. SSRF-guarded — refuses private/loopback/internal hosts. Returns url, title, format, length, content (and finalUrl when redirected, plus up to 50 extracted links when links=true), with a meta envelope. Works with no API key.

輸入參數url必填linksformatmaxChars

版本歷史

1
  1. v1.0.0最新Sep 16, 2026

評分

已滿足適用檢查項的 90%

該評分衡量此收錄條目的完整程度,依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價,也不表示伺服器安全或目前可用:可連通性是帶日期的檢查結果,評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分
  • 已宣告認證方式必需10 分
  • 已列出工具必需15 分
  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分

    請在儲存庫或註冊中心條目中宣告授權條款。