DeepSift Search

io.github.pavanmulka由 GitHub 账号 pavanmulka 发布v0.1.0更新于 Nov 5, 2026

Live Google web search for AI agents that returns only the passages that answer the question.

已验证Streamable HTTP可网页运行网页搜索与抓取知识与记忆

概览

AI 生成的概览

托管式网页搜索 MCP 服务器，执行 Google 搜索并只返回能回答问题的网页段落，另提供读取、站点地图、爬取与深度研究工具。

可以做什么
  • 针对一个主题搜索网页，一次回复中给出多个网站中最相关的段落
  • 读取最多 20 个网页或 PDF 的全文，或只取与指定主题相关的部分
  • 列出某个站点的页面地址，并爬取同一站点的多个页面、保留相关内容
  • 执行多轮搜索的调研任务，返回带编号来源的书面报告
  • 提供与 Tavily 相同的 search、extract、crawl、map 请求与响应格式
适用场景
适合需要带来源的实时网页答案、而不只是链接列表的场景，也适合希望使用托管端点而非本地自建搜索流程的用户。可用于调研、价格或文档查询以及整站阅读。
运行要求
远程 streamable HTTP 端点，无需本地安装。需要一个 Apify API 令牌，以 Bearer 形式放在 Authorization 请求头中。免费 Apify 账户每月含一定额度，用量按搜索次数和读取页数计费。
安装前请注意
按使用付费：每 1,000 次搜索 4 美元、每 1,000 个页面 1 美元，从 Apify 额度中扣除。Apify 令牌属于机密，可访问你的账户。查询内容和抓取的页面内容会发送到托管服务及 Google。调研任务可能耗时数分钟。

安装

在 SourceWeft 中

  1. 打开 控制台中的 DeepSift Search，将其添加到工作区。
  2. 连接 Authorization token 凭据。它们会被私密存储，绝不会在公开页面上展示。
  3. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体，或 SourceWeft 对话中。智能体会读取该服务器的安装指南，先向你展示发布者、端点和登录方式，经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.pavanmulka%2Fdeepsift-search/install.md，并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前，先告诉我它的发布者、端点和登录方式，等我确认。不要把 API 密钥写进命令或我会提交的文件；问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "deepsift-search": {
      "type": "http",
      "url": "https://aadstra--deepsift-search.apify.actor/mcp",
      "headers": {
        "Authorization": "Bearer <YOUR_TOKEN>"
      }
    }
  }
}

README

DeepSift

A free, self-hosted alternative to Tavily for Claude Code. Ask one question; DeepSift searches the web, reads every page, and hands back only the paragraphs that matter. It runs on your laptop and uses the Claude subscription you already pay for.

Tavily is a paid web search API built for AI agents. You send one question; it searches the web, opens the best pages, cuts out the passages that answer the question, and sends it all back in one tidy reply. That's great, but every call costs credits.

DeepSift does the same job on your own computer, in three steps:

  1. Find the right websites, using Claude's web search on your existing Claude subscription (with free DuckDuckGo as a backup).
  2. Read every page itself: HTML and PDFs, with menus, ads and clutter stripped out.
  3. Pick the passages that best match your question, and give them back in one reply.

No per-call bill, no API key to buy.

Don't want to run it yourself? Use DeepSift Search on Apify

[DeepSift Search icon]

DeepSift Search is the hosted version: real Google results, the same passage picking, nothing to install, and it works for apps and teams, not just your own laptop.

  • Tavily drop-in: same /search, /extract, /crawl and /map requests and responses. Change one line:

    python
    from tavily import TavilyClientclient = TavilyClient(api_key="<YOUR_APIFY_TOKEN>", api_base_url="https://aadstra--deepsift-search.apify.actor")

  • MCP server for Claude, Cursor and others: https://aadstra--deepsift-search.apify.actor/mcp with your Apify token as a Bearer header.

  • Claude Code plugin: /plugin marketplace add pavanmulka/deepsift, then /plugin install deepsift-search@deepsift. Claude Code asks for your Apify token once.

  • Pay per use, no subscription: $4 per 1,000 searches, $1 per 1,000 pages read. Failed requests are free, and a free Apify account includes $5 of credit every month.

The Apify version lives in .actor/ and deepsift/actor.py; it never uses your Claude subscription.

What Claude Code gets

After installing, every Claude Code session has these tools:

ToolWhat you'd askWhat it does
web_search"web_search Databricks Unity Catalog pricing"One topic → many websites → the best passages from each, in one reply
web_read"read these 3 pages and focus on pricing"Full text of up to 20 pages or PDFs, or only the parts about a topic
web_map"list all pages on docs.example.com"A site's page addresses
web_crawl"read the pricing pages on example.com"Reads many pages of one site, keeps the relevant ones
web_research"research the top 5 vector databases with sources"A multi-search investigation that returns a written report with numbered sources (takes minutes)

The same abilities work from any terminal:

sh
deepsift search "Databricks Unity Catalog pricing" --advanceddeepsift read https://example.com/page --focus "pricing"deepsift map https://docs.example.comdeepsift crawl https://docs.example.com --instructions "pricing"deepsift research "Which vector databases support hybrid search?"

Install (about 1 minute)

You need a Mac or Linux computer with Claude Code and Python 3.10+.

sh
git clone https://github.com/pavanmulka/deepsift.gitcd deepsift./install.shclaude auth login        # only if you aren't signed in to Claude Code's command line yet

Open a new Claude Code session and ask it to use web_search. That's it.

./uninstall.sh removes it from Claude Code again.

How it compares to Tavily

TavilyDeepSift
CostPay per creditFree (uses your Claude plan)
Speed per search1–2 s~10–30 s through Claude; ~1–3 s on the DuckDuckGo backup
Search, extract, crawl, map, research✓✓
Same request/response format—✓ (optional local HTTP API, see below)
Runs whereTheir serversYour laptop

Limits, honestly:

  • Searches through Claude count toward your Claude plan's usage limits (3 at a time at most). Repeat searches within 24 h come from a local cache.
  • DuckDuckGo pauses you after a burst of quick searches; the tool then rests it for 15 minutes.
  • Pages that only show their text after JavaScript runs come back thin.
  • It's meant for your own personal use with your own Claude subscription. To power an app for other people, use a proper paid API.

Optional: a local Tavily-compatible API

./run.sh starts an HTTP server on 127.0.0.1:8787 with Tavily's paths and fields (/search, /extract, /crawl, /map, /research, GET /research/{id}, /usage). Code written for Tavily can point its base URL there. Every reply reports usage.credits: 0.

Settings

Copy .env.example to .env to change engines, models, cache time, or set an API key for the HTTP server.

How it works

question ─► find sites (Claude WebSearch via `claude -p`, or DuckDuckGo)         ─► download pages in parallel (httpx) ─► readable text (trafilatura, pypdf)         ─► split into passages ─► rank by relevance to the question (BM25) ─► one reply

claude -p always runs on your subscription sign-in. ANTHROPIC_API_KEY is removed before it starts, so it never bills an API account by accident.

Tests

sh
.venv/bin/python -m pytest -q tests

The tests run offline, using a stand-in for claude.

License

MIT

来源：README.md，提交 b37287a

工具

0
该服务器需要 API 密钥等凭据，因此 SourceWeft 无法在没有凭据的情况下读取其工具，也未收录。

版本历史

1
  1. v0.1.0最新Nov 5, 2026

评分

已满足适用检查项的 85%

该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list，或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分