PaperPilot AI — MCP-Powered Semantic Research Platform

An MCP (Model Context Protocol) server that lets an AI client search arXiv, fetch papers, and answer questions grounded in the actual paper text — a full retrieve-and-generate RAG pipeline exposed through all three MCP primitives (tools, resources, and prompts), plus a standalone demo UI.

Demo

Watch the 1-minute Demo

Search & Fetch

Ask a Paper

Compare Papers

Why this exists

Most "I built an MCP server" projects stop at wrapping a single API call in a tool. This project instead demonstrates a complete, real pipeline:

search → fetch → chunk → embed → retrieve → generate

...exposed through MCP so it's usable directly from Claude Desktop or any other MCP client, with a companion Streamlit UI for visual demos.

Architecture

flowchart LR
    A[User question] --> B[search_arxiv]
    B --> C[fetch_paper]
    C --> D[Chunk text]
    D --> E[Embed chunks<br/>sentence-transformers]
    E --> F[(SQLite<br/>persisted cache)]
    F --> G[ask_paper]
    G --> H[Embed question]
    H --> I[Cosine similarity<br/>retrieve top-k chunks]
    I --> J[Groq LLM<br/>generate grounded answer]
    J --> K[Answer + source excerpts]

Loading

Features

MCP Primitive	Name	What it does
Tool	search_arxiv	Search arXiv by keyword
Tool	fetch_paper	Download a paper's PDF, extract + chunk + embed its text
Tool	ask_paper	Answer a question grounded in a fetched paper's content (RAG)
Tool	compare_fetched_papers	Compare two fetched papers' methods and contributions
Resource	papers://list	Browse every paper fetched so far
Resource	papers://{paper_id}	View a specific paper's full extracted text
Prompt	literature_review	Scaffolds a multi-paper research workflow
Prompt	compare_papers	Scaffolds a structured two-paper comparison

Plus:

• Persistence — fetched papers survive a server restart (SQLite), not just in-memory.
• Error handling — bad paper IDs, network failures, and LLM errors return clean messages instead of crashing.
• 20 automated tests — all external calls (arXiv, PDF download, Groq) are mocked, so the suite runs in seconds with zero API cost. CI runs them on every push.
• Demo UI — a Streamlit app that reuses the exact same functions as the MCP server (no duplicated logic), with retrieved excerpts shown visually to make the RAG mechanism transparent.

Tech stack

• Protocol: MCP via FastMCP
• Embeddings: sentence-transformers (all-MiniLM-L6-v2, runs locally, no API cost)
• LLM: Groq (llama-3.3-70b-versatile)
• Data: arXiv API, pypdf for text extraction
• Storage: SQLite
• Testing: pytest + pytest-mock
• Demo UI: Streamlit

Setup

1. Clone and install

git clone https://github.com/ayushisingh51/PaperPilot-AI.git
cd PaperPilot-AI
python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Add your API key

cp .env.example .env
# then edit .env and add your free key from https://console.groq.com/keys

3. Run it

As an MCP server (test in the MCP Inspector):

fastmcp dev server.py

Connected to Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "research-assistant": {
      "command": "/absolute/path/to/venv/bin/python",
      "args": ["/absolute/path/to/research-mcp/server.py"]
    }
  }
}

As a standalone demo UI:

streamlit run demo_app.py

4. Run the tests

pytest tests/ -v

Project structure

PaperPilot-AI/
├── server.py                # MCP server: tools, resources, prompts
├── demo_app.py               # Standalone Streamlit demo UI
├── tests/
│   ├── conftest.py           # Test fixtures (temp DB, dummy API key)
│   └── test_server.py        # 20 tests, all external calls mocked
├── .github/workflows/
│   └── tests.yml              # CI: runs tests on every push
├── requirements.txt
├── .env.example
└── LICENSE

Possible extensions

• Swap the naive top-k retrieval for a proper vector DB (Chroma/FAISS) as the paper library grows.
• Add streaming responses for the generation step.
• Support multi-paper synthesis in a single ask call instead of one at a time.

License

MIT