Open Deep Research

The Open Deep Research MCP Server iteratively explores a topic, scrapes relevant information, and utilizes AI to compile a comprehensive report.

You will get a full markdown file with all the findings, the sources it used, and even a reliability score for each source.

Features

• 🔍 Deep Iterative Research – Generates targeted search queries and performs multi-level investigation
• 📊 Source Reliability Scoring – Evaluates sources with detailed 0-1 scoring and reasoning
• 🎯 Configurable Scope Control – Adjusts research depth and breadth through parameters
• 📝 Comprehensive Reports – Produces detailed markdown reports with findings and source metadata
• 🔄 Follow-up Question Generation – Creates targeted questions to refine research direction
• ⚡ Dual Operation Modes – Functions as MCP server or standalone CLI tool
• 🌐 Local Firecrawl Support – Option to use local Firecrawl instance instead of API
• 📈 Observability Integration – Optional Langfuse tracking for research flow monitoring

Use Cases

• Academic Research Projects – Conduct literature reviews and gather citations with reliability assessments for thesis work or publication research
• Market Analysis – Investigate industry trends, competitor analysis, and market conditions with structured documentation and source verification
• Technical Documentation – Research best practices, implementation patterns, and technology comparisons for development projects
• Due Diligence Research – Perform comprehensive background research on companies, technologies, or individuals with source reliability tracking

How To Use It

1. Clone the package from GitHub and install the necessary dependencies.

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

2. Create a local environment file.

cp .env.example .env.local

You can open .env.local to see the available configuration options, but you don’t need to change anything for the basic setup.

3. Build the project.

npm run build

4. You have two ways to run it: as a CLI tool or as an MCP server.

• CLI Version: To run a one-off research task directly from your terminal, use this command:

npm run start "Your research query here"

• MCP Server Version: To run it as a persistent server that AI clients like Claude Desktop can connect to, use this command:

npm run start:http

This starts the server on http://localhost:3000/mcp. It operates without session management. You can then add this URL to your MCP client.

Advanced Setup

Using a Local Firecrawl Instance (Free Option)

The server uses Firecrawl for scraping. To avoid using their API key and any associated costs, you can run Firecrawl locally. The developer even has a fork that uses searXNG as the search backend, which means you don’t need a search API key either.

1. Set up local Firecrawl:

git clone https://github.com/Ozamatash/localfirecrawl
cd localfirecrawl
# Follow the setup instructions in the localfirecrawl README

2. Update your .env.local file in the deep-research directory to point to your local Firecrawl instance:

FIRECRAWL_BASE_URL="http://localhost:3002"

Observability with Langfuse

If you want to track and visualize the research flows, you can connect it to Langfuse. Add your Langfuse keys to the .env.local file:

# Add to .env.local
LANGFUSE_PUBLIC_KEY="your_langfuse_public_key"
LANGFUSE_SECRET_KEY="your_langfuse_secret_key"

The server works perfectly fine without these keys if you don’t need observability.

FAQs

Q: How does the source reliability scoring work?
A: The system evaluates each source using multiple criteria including domain authority, content quality, publication date, and citation patterns. Sources receive scores from 0-1 with detailed reasoning. High-reliability sources (≥0.7) are prioritized, while lower-scoring sources undergo additional verification.

Q: What’s the difference between the MCP version and CLI version?
A: The CLI version includes interactive follow-up questions to refine research direction, while the MCP version focuses on direct research execution within AI agent workflows. Both generate the same quality reports with source reliability assessments.

Q: Can I use this without API keys for web scraping?
A: Yes, you can set up a local Firecrawl instance using the provided fork that includes searXNG as the search backend. This eliminates the need for external API keys while maintaining full functionality.

Q: How does the iterative research process work?
A: The system generates initial search queries based on your input, evaluates results, identifies knowledge gaps, then creates new targeted queries for deeper investigation. This continues until the specified depth limit is reached or research goals are satisfied.The server generates comprehensive markdown reports including structured findings, source metadata with reliability scores, research methodology notes, and suggested follow-up directions. Reports can be exported or integrated into documentation workflows.

Q: Is observability tracking required?
A: No, observability with Langfuse is optional. The system functions normally without tracking keys and only enables enhanced monitoring when configured.