Deep Research

A simple, direct, and highly scalable deep research tool based on multi-agent architecture, supporting arbitrary LLM and MCP tool integration.

✨ Features

• 🔌 OpenAI API Compatible: Supports any LLM compatible with OpenAI API, such as OpenAI, OpenRouter, DeepSeek, etc., and does not require the model to have tool-calling capabilities.
• 🛠️ Universal MCP Extension: Supports integration of arbitrary MCP tools (stdio, streamable or sse) to extend Agent capabilities.
• 🌐 Simple and Intuitive: Exposes both HTTP and MCP interfaces with clean APIs for easy integration.
• ⚡ High-Performance Async: Built on FastAPI, supports high-concurrency request processing.

How It Works

flowchart TD
    Task(["Task"]) --> AgentSystem

    subgraph AgentSystem["Loop"]
        Planner["📋 Planner Agent"]
        Worker1["🔧 Worker  Agent 1"]
        Worker2["🔧 Worker  Agent 2"]
        Worker3["🔧 Worker  Agent N"]

        Planner --> Worker1
        Planner --> Worker2
        Planner --> Worker3

        Worker1 --> Planner
        Worker2 --> Planner
        Worker3 --> Planner

        Reporter["📝 Reporter Agent"]
        Planner --> Reporter
    end

    Reporter --> Report(["Report"])

The general workflow is as follows:

1. The user submits a research task to the system.

2. The planner analyzes the task and dispatches initial-stage subtasks to workers(config no more than max_subtasks subtasks).

3. Workers execute the subtasks in parallel, and return the subtask reports.

4. Subtasks reports are aggregated and passed to the planner.

5. The planner analyzes the context and the subtask results to determine next action.

6. Repeat this process until one of the following conditions is met:

   • The Planner proactively invokes the Reporter to generate a report when sufficient information is deemed available..
   • The number of iterations reaches the max_reasoning_times configuration, at which point the system forcibly passes the existing information to the Reporter and requests it to generate a report.

7. The system delivers the final report to the user.

🚀 Quick Start

System Requirements

• Python 3.10+

1. Clone the Repository

git clone https://github.com/troyhantech/deep-research.git
cd deep-research

2. Install Dependencies

pip install uv
uv pip install -r requirements.txt

# or

pip install -r requirements.txt

3. Configuration

3.1 Environment Variables

Copy the template file:

cp .env.example .env

Edit the .env file and configure your API keys:

OPENAI_API_KEY="your-openai-api-key"
OPENAI_BASE_URL="https://api.openai.com/v1/"

# Optional: LangSmith tracing
LANGSMITH_TRACING="true"
LANGSMITH_ENDPOINT="https://api.smith.langchain.com"
LANGSMITH_API_KEY="your-langsmith-api-key"
LANGSMITH_PROJECT="your-langsmith-project"

3.2 Application Configuration

Copy the template file:

cp config.toml.example config.toml

Edit config.toml to configure agents and MCP services:

[agents]
[agents.planner]
model = "gpt-4o"
max_reasoning_times = 5
max_tokens = 4096
max_subtasks = 10

[agents.reporter]
model = "gpt-4o"
max_tokens = 4096

[agents.worker]
model = "gpt-4o"
max_tokens = 4096
max_reasoning_times = 5

# Support three standard MCP transport as worker tools: streamable_http、stdio or sse.
[mcp_servers]

# use stdio transport
[mcp_servers.tavily_stdio]
enabled = false
type = "stdio"
command = "npx"
args = ["-y", "mcp-remote", "https://mcp.tavily.com/mcp/?tavilyApiKey=your-tavily-api-key"]
include_tools = ["tavily_search"] # Only provide the tavily_search tool to Worker, add only the tools that need to be used to save context usage. If left empty, all tools will be provided to Worker by default.

# use streamable_http transport
[mcp_servers.tavily_streamable_http]
enabled = true
type = "streamable_http"
url = "https://mcp.tavily.com/mcp/?tavilyApiKey=your-tavily-api-key"
include_tools = ["tavily_search"] # if empty, default all

# or sse transport (Deprecated - recommend using streamable_http instead)
[mcp_servers.sse_server_example]
enabled = false
type = "sse"
url = "sse_server_url"
include_tools = ["tavily_search"]

4. Start the Service

The command to start the service is python main.py, with the default launch mode being mcp_stdio. You can modify the launch mode by specifying the mode parameter:

python main.py

options:
  --env-file path to .env file, default: ./.env
  --config-file path to config.toml, default: ./config.toml
  --mode Launch mode: mcp_stdio, mcp_streamable_http, or http_api, default: mcp_stdio
  --host default: 0.0.0.0
  --port default: 8000

4.1 Start MCP STDIO Service

The command to start the service in MCP STDIO mode is:

python main.py --mode mcp_stdio

To launch via MCP client, you need to explicitly specify the absolute paths to configuration files. For example, you can configure in Claude Desktop:

{
  "mcpServers": {
    "deep-research": {
      "command": "/ABSOLUTE/PATH/TO/python", // The path to python, you can use `which python` on Linux or Mac, or `where python` on Windows.
      "args": [
        "/ABSOLUTE/PATH/TO/main.py", // The absolute path to the project's main.py on your machine.
        "--env-file",
        "/ABSOLUTE/PATH/TO/.env", // The absolute path to the project's .env file on your machine.
        "--config-file",
        "/ABSOLUTE/PATH/TO/config.toml", // The absolute path to the project's config.toml file on your machine.
        "--mode",
        "mcp_stdio"
      ]
    }
  }
}

4.2 Start MCP STREAMABLE HTTP Service

The command to start the service in MCP STREAMABLE HTTP mode is:

python main.py --mode mcp_streamable_http --host 0.0.0.0 --port 8000

Now you can remotely access your deep-research service by configuring http://localhost:8000/mcp/ in the MCP client.

4.3 Start HTTP API Service

The command to start the service in HTTP API mode is:

python main.py --mode http_api --host 0.0.0.0 --port 8000

Now you can access via:

• API endpoint: http://localhost:8000/deep-research
• Web interface: http://localhost:8000/web

Send your research task via HTTP POST request to http://localhost:8000/deep-research, and after some time, you'll receive a research report.

Request Body:

{
  "task": "Analyze Bitcoin price trends for the next month, output in English"
}

Response Body:

{
  "result": "# Bitcoin (BTC) Price Trend Analysis for Next Month\n\n## Introduction\n\nThis report aims to analyze the core driving factors of Bitcoin (BTC) price trends for the next month (August 18, 2025 to September 17, 2025). By focusing on market liquidity, technical charts, and key macroeconomic signals, we strive to identify the main market contradictions and determine BTC's potential direction and key price levels during this period..."
}

Quick Start Examples:

• You can quickly test via browser by visiting http://localhost:8000/web
• You can request the API using the following command:

curl -X POST "http://localhost:8000/deep-research" \
     -H "Content-Type: application/json" \
     -d '{"task": "Analyze Bitcoin price trends for the next month, output in English"}'

{
  "task": "Analyze Bitcoin price trends for the next month, output in English",
  // Optional, defaults to agents config in config.toml
  "config": {
    "planner": {
      "model": "gpt-4o",
      "max_reasoning_times": 5,
      "max_tokens": 4096,
      "max_subtasks": 10
    },
    "worker": {
      "model": "gpt-4o",
      "max_tokens": 4096,
      "max_reasoning_times": 5
    },
    "reporter": {
      "model": "gpt-4o",
      "max_tokens": 4096
    }
  }
}

Research Task Tips

Research tasks should be clear and specific. A well-defined research task should include:

1. Research topic: What is the subject of the research?
2. Research background: Why is this research being conducted? What are the objectives?
3. Research requirements: What are the specific requirements for the research?

❓ Frequently Asked Questions (FAQ)

Q: Which LLM models are supported?

A: Supports any model compatible with OpenAI API, including OpenAI GPT series, OpenRouter, etc.

Q: Must I use models with Function Call capabilities?

A: No, this project does not depend on the model's Function Call functionality. Any large language model will work.

Q: Is prompt caching supported?

A: No, prompt caching is not supported. However, most of the models support automatic caching mechanism, and there are enough models support it.

Q: How to add custom MCP tools?

A: Add your MCP service configuration in the [mcp_servers] section of the config.toml file. For example: integrate custom knowledge base search MCP services.

Q: If I don't add any MCP tools, will there be any impact?

A: No, there will be no impact. The system will still work, but agent won't be able to use MCP tools, and answer using their existing knowledge.

Q: Only Python 3.10+ is supported?

A: No, the author only tests with Python 3.10+, but it should work with 3.8+ as well.

💡 Short-term Plan

• Implement streaming output of the execution process and report for faster response.
• Output the consumed token data.
• Combine echarts to generate interactive charts, making the report more visually engaging.

License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Contributing

We welcome all forms of contributions! Whether it's reporting bugs, suggesting new features, or submitting code improvements.

🐛 Issue Reporting

If you find bugs or have feature suggestions, please submit them on the Issues page.

⭐ Star History

If this project helps you, please give us a ⭐!

Star History Chart