🎓 ArXiv Research Assistant (MCP Server)

A Research Agent server built with the Model Context Protocol (MCP). Search ArXiv, extract structured findings from papers, build a bibliography, and discover insights (keyword trends, topic clusters, a research timeline, an AI-written literature review) — all as tools an AI assistant can call directly, plus a full multi-page web dashboard with editing built in. Runs either as a local server launched by your IDE, or hosted remotely.

🚀 Features

• Automated Search: Query ArXiv for papers on any topic.

• Intelligent Extraction: heuristic keyword-based extraction (free), or LLM-based extraction via Claude Haiku 4.5 (higher accuracy, small API cost) — of the core "Problem", "Method", and "Result" from abstracts.

• Bibliography: saved to a Postgres database (dedupes automatically by paper URL), fully editable — update or delete any saved paper.

• Data Science Pipeline — all available as MCP tools, so an agent can trigger them directly:

  • Visualize Trends: bar chart of dominant research methods.

  • Research Timeline: stacked bar chart of how your saved papers' topics have shifted year over year.

  • Topic Modeling: NMF (Non-Negative Matrix Factorization) discovers hidden research themes, then a single Claude Haiku call turns the raw keyword clusters into short readable labels — labels can also be renamed manually at any time.

  • Literature Review: one Claude Haiku call synthesizes the whole bibliography into a cohesive, numbered-citation narrative review — editable directly, or regenerated from scratch.

• Web Dashboard: a real multi-page site (Overview / Papers / Topics / Review) at the deployed root URL, gated behind a login page — not open to the whole internet.

Related MCP server: arXiv Research MCP Server

Installation

Prerequisites

• Python 3.10+

• A Postgres database — Neon has a free tier and is what this was built/tested against.

• An Anthropic API key for the LLM-based tools (extract_key_findings_llm, the topic-labeling step inside discover_research_topics, and generate_literature_review). Not needed for the rest of the server.

Setup

1. Clone the repository:

   git clone <your-repo-url>
   cd arxiv-research-mcp

2. Create a Virtual Environment:

   python3 -m venv .venv
   source .venv/bin/activate

3. Install Dependencies:

   pip install -r requirements.txt

4. Configure environment variables — copy .env.example to .env and fill in DATABASE_URL (your Neon connection string) and ANTHROPIC_API_KEY. Leave MCP_SERVER_URL/MCP_AUTH_TOKEN unset for local use.

5. Initialize the database schema (one-time):

   python db.py

⚙️ Configuration — Local (stdio)

Add the server to your IDE's MCP settings (e.g., mcp-servers.json in VS Code or Claude Desktop):

{
  "mcpServers": {
    "research-assistant": {
      "command": "/absolute/path/to/your/arxiv-research-mcp/.venv/bin/python",
      "args": [
        "/absolute/path/to/your/arxiv-research-mcp/research_server.py"
      ],
      "env": {
        "PYTHONPATH": "/absolute/path/to/your/arxiv-research-mcp",
        "DATABASE_URL": "postgresql://...",
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Note: Replace the absolute paths with the actual full paths on your machine.

☁️ Configuration — Remote (hosted on Vercel)

This same server can run as a remote MCP endpoint instead of a process your IDE launches. It authenticates over OAuth rather than a static bearer token — clients like Claude.ai's "Add custom connector" only speak full OAuth for connectors that declare auth, so this server implements a minimal single-user authorization server (dynamic client registration + a login form) rather than expecting the client to accept a pasted token directly.

1. Deploy this repo to Vercel (it auto-detects the Python entrypoint index.py).

2. Set these Vercel environment variables:

   • DATABASE_URL, ANTHROPIC_API_KEY (same as local)

   • MCP_SERVER_URL — the public URL this deployment is reachable at, e.g. https://<your-project>.vercel.app

   • MCP_AUTH_TOKEN — the password you'll type into the login form during setup (not something you paste into the client). Generate one with:

     python -c "import secrets; print(secrets.token_urlsafe(32))"

3. In your MCP client (e.g. Claude.ai → Settings → Connectors → Add custom connector), enter the server URL as https://<your-project>.vercel.app/mcp. The client will register itself and redirect you to a login page hosted by this server — enter the MCP_AUTH_TOKEN there to authorize it.

OAuth (and with it, the login gate) only activates when MCP_SERVER_URL is set. Leaving it unset — the local-dev default — means the deployed endpoint would be wide open, letting anyone who finds the URL call your tools and spend your Anthropic budget. Set both MCP_SERVER_URL and MCP_AUTH_TOKEN together for a real hosted deployment.

💡 Usage

Everything below is available as an MCP tool an agent can call directly — no separate terminal step needed. Just ask your AI Assistant things like:

"Find 5 papers on 'Reinforcement Learning from Human Feedback', extract their key findings, and save them to my bibliography."

"Now visualize the keyword trends in my bibliography, tell me the dominant research topics, and write me a literature review."

MCP Tools

• search_arxiv(query, max_results) — Search ArXiv for papers.

• extract_key_findings(abstract) — Free, heuristic Problem/Method/Result extraction.

• extract_key_findings_llm(abstract) — Same extraction via Claude Haiku 4.5 — more accurate, small API cost.

• save_to_bibliography(paper_metadata) — Upsert a paper (with its findings) into the database, keyed by its PDF URL.

• update_paper(paper_id, ...) — Edit any field of a saved paper (title, authors, published, problem, method, result). Fields left unset are unchanged.

• delete_paper(paper_id) — Permanently remove a paper from the bibliography.

• visualize_keyword_trends() — Bar chart of common method keywords, returned as an inline image.

• visualize_research_trends() — Stacked bar chart of papers per topic per year, returned as an inline image.

• discover_research_topics(num_topics) — NMF topic modeling over the bibliography, plus Claude-generated readable labels for each topic.

• rename_topic(topic_id, label) — Manually rename a topic's label; applies to every paper sharing that topic, no recompute needed.

• generate_literature_review() — Synthesizes the whole bibliography into one narrative review with numbered [n] citations, via a single Claude Haiku call.

• update_literature_review(content) — Manually overwrite the saved review's text without regenerating it.

Web Dashboard

Once deployed (or run locally), visit the server's root URL in a browser — it's gated behind a login form using the same MCP_AUTH_TOKEN as the OAuth setup.

• Overview (/) — stat tiles, both charts, a recently-saved papers preview.

• Papers (/dashboard/papers) — the full bibliography with search/sort, a Problem preview per row, and Edit/Delete actions. Each paper's edit page (/dashboard/papers/{id}/edit) exposes every field, including the full Problem/Method/Result text.

• Topics (/dashboard/topics) — every discovered topic with its raw NMF keywords and paper count, an inline rename form per topic, and a Recompute Topics button.

• Review (/dashboard/review) — the literature review in an editable textarea, with Save and Regenerate (regenerating warns first if you have unsaved manual edits).

Running the analysis scripts standalone

Each analysis tool is also a runnable CLI script, if you'd rather generate insights from the terminal directly (writes its output to a local file, for convenience — the MCP tools above never touch disk):

python analyze_references.py     # -> method_keywords.png
python research_trends.py        # -> research_trends.png
python topic_modeling.py         # updates topic assignments in the database

📂 Project Structure

• research_server.py: The core MCP server — defines every tool listed above.

• index.py: Vercel entrypoint — exposes the same server over Streamable HTTP.

• db.py: Postgres access (schema, paper CRUD, topic updates, literature review cache).

• oauth_provider.py: Minimal single-user OAuth authorization server for the hosted deployment (dynamic client registration + a login form gated by MCP_AUTH_TOKEN).

• dashboard.py / dashboard_templates.py: The web dashboard's routes/CRUD logic and its Jinja2 templates, with their own login-cookie gate (also keyed by MCP_AUTH_TOKEN, separate from the OAuth flow above).

• llm.py: Claude Haiku 4.5 calls (structured-output extraction, topic labeling, literature review generation).

• chart_style.py: Shared matplotlib styling so every chart looks consistent.

• analyze_references.py: Keyword-frequency visualization (also the visualize_keyword_trends tool).

• research_trends.py: Topic-over-time analysis (also the visualize_research_trends tool).

• topic_modeling.py: Scikit-learn NMF topic modeling (also the discover_research_topics tool).

• requirements.txt: Python dependencies.

Built with mcp, arxiv, anthropic, pandas, scikit-learn, Jinja2, and Neon Postgres.