MCP Integration

Omelette exposes its capabilities via the Model Context Protocol (MCP), allowing AI coding assistants like Claude Code, Codex, and Cursor to interact with your scientific knowledge bases.

Overview

The MCP server runs as an ASGI sub-application mounted at /mcp on the Omelette backend. It provides:

• Tools: Executable actions (search, lookup, add papers)
• Resources: Read-only data access (knowledge bases, papers, chunks)
• Prompts: Pre-built prompt templates for common tasks

Available Tools

Tool	Description
list_knowledge_bases	List all knowledge bases with paper counts
search_knowledge_base	RAG-powered semantic search within a knowledge base
lookup_paper	Find paper by DOI or title (local DB + Crossref)
find_citations	Find citation candidates for a given text passage
add_paper_by_doi	Add a paper to a knowledge base by DOI
get_paper_summary	Get paper metadata and abstract
search_papers_by_keyword	Multi-source federated keyword search

Available Resources

URI Pattern	Description
omelette://knowledge-bases	List all knowledge bases
omelette://knowledge-bases/{id}	Knowledge base details with paper list
omelette://papers/{id}	Paper metadata
omelette://papers/{id}/chunks	Paper text chunks

Connecting from AI IDEs

Claude Code / Codex

Add to your MCP configuration:

{
  "mcpServers": {
    "omelette": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "omelette": {
      "url": "http://localhost:8000/mcp"
    }
  }
}

Stdio Mode

For local development without HTTP:

cd backend
python -m app.mcp_server

Example Usage

Once connected, you can ask your AI assistant:

• "Search the ML knowledge base for papers about transformer attention"
• "Find citations for this paragraph: ..."
• "Add the paper with DOI 10.1234/example to my knowledge base"
• "Summarize paper #42"