这个仓库是对代理编码的实验。95%的代码由LLM编写

打开深度研究MCP服务器

一款由人工智能驱动的研究助手，能够对任何主题进行深入、迭代的研究。它结合了搜索引擎、网页爬虫和人工智能，深入探讨主题并生成全面的报告。可作为模型上下文协议（MCP）工具或独立的CLI使用。查看 exampleout.md，了解报告可能的样貌。

快速启动

1. 克隆与安装：

2. 在.env.local中建立环境：

3. 构建：

4. 运行CLI版本：

5. 用 Claude Desktop 测试 MCP 服务器： 请按照服务器快速入门底部的指南，将服务器添加到Claude Desktop： https://modelcontextprotocol.io/quickstart/server

对于远程服务器：可流式HTTP

服务器运行在 http://localhost:3000/mcp 上，无需会话管理。

特色

• 通过生成有针对性的搜索查询进行深入、迭代的研究
• 控制研究范围的深度（多深）和广度（多广）参数
• 通过详细评分（0-1）和推理评估源可靠性
• 优先考虑高可靠性来源（≥0.7），并验证较不可靠的信息

DISCLAIMER

This repo is an experiment on agent coding. 95% of the code is written by LLM's 

Open Deep Research MCP Server

An AI-powered research assistant that performs deep, iterative research on any topic. It combines search engines, web scraping, and AI to explore topics in depth and generate comprehensive reports. Available as a Model Context Protocol (MCP) tool or standalone CLI. Look at exampleout.md to see what a report might look like.

Quick Start

1. Clone and install:

git clone https://github.com/Ozamatash/deep-research
cd deep-research
npm install

2. Set up environment in .env.local:

# Copy the example environment file
cp .env.example .env.local

3. Build:

# Build the server
npm run build

4. Run the cli version:

npm run start

5. Test MCP Server with Claude Desktop:
   Follow the guide thats at the bottom of server quickstart to add the server to Claude Desktop:
   https://modelcontextprotocol.io/quickstart/server

For remote servers: Streamable HTTP

npm run start:http

Server runs on http://localhost:3000/mcp without session management.

Features

• Performs deep, iterative research by generating targeted search queries
• Controls research scope with depth (how deep) and breadth (how wide) parameters
• Evaluates source reliability with detailed scoring (0-1) and reasoning
• Prioritizes high-reliability sources (≥0.7) and verifies less reliable information
• Generates follow-up questions to better understand research needs
• Produces detailed markdown reports with findings, sources, and reliability assessments
• Available as a Model Context Protocol (MCP) tool for AI agents
• For now MCP version doesn’t ask follow up questions
• Natural-language source preferences (avoid listicles, forums, affiliate reviews, specific domains)

Model Selection (OpenAI, Anthropic, Google, xAI)

Pick a provider and model per run.

• CLI: you will be prompted for provider and model. Example: openai + gpt-5.2.
• MCP/HTTP: pass model, e.g. openai:gpt-5.2 (also accepts openai/gpt-5.2).

Set the corresponding API key in .env.local:

OPENAI_API_KEY=...
ANTHROPIC_API_KEY=...
GOOGLE_API_KEY=...
XAI_API_KEY=...

Optionally set default models per provider:

OPENAI_MODEL=gpt-5.2
ANTHROPIC_MODEL=claude-opus-4-5
GOOGLE_MODEL=gemini-3-pro-preview
XAI_MODEL=grok-4-1-fast-reasoning

If you use a non-default OpenAI endpoint:

OPENAI_ENDPOINT=https://api.openai.com/v1

How It Works

flowchart TB
    subgraph Input
        Q[User Query]
        B[Breadth Parameter]
        D[Depth Parameter]
        FQ[Feedback Questions]
    end

    subgraph Research[Deep Research]
        direction TB
        SQ[Generate SERP Queries]
        SR[Search]
        RE[Source Reliability Evaluation]
        PR[Process Results]
    end

    subgraph Results[Research Output]
        direction TB
        L((Learnings with
        Reliability Scores))