fro-wang-academic-tools-mcp 0.1.4

All-in-one MCP server for academic paper processing: OCR, metadata, structure extraction, translation, summary, arXiv, and Zotero.

fro-wang-academic-tools-mcp

fro-wang-academic-tools-mcp is an MCP server for academic paper workflows, built on FastMCP. It combines OCR, metadata extraction/enrichment, section parsing, translation, summary generation, arXiv tools, and Zotero integration.

Detailed system design is documented in docs/README_SYSTEM.md.

Motivation

This project is dedicated to managing academic papers locally to empower AI Agents. While AI Agents excel at local tasks, they struggle with the dominant format of academic literature: PDFs. When you are conducting literature reviews, reading, or searching for papers, most of your local library consists of PDF files.

Local coding agents prefer plain text. They rely on tools like grep or semantic search to navigate local files. These tools cannot natively search inside PDFs, making it difficult for agents to assist with literature management effectively.

Having frequently encountered this bottleneck, I built fro-wang-academic-tools-mcp to bridge the gap. These tools extract the core workflow from my website, frowang.com, converting PDFs into agent-friendly formats (like Markdown) and enriching them with metadata.

Features

• End-to-end paper pipeline via process_paper
• OCR with MinerU (ocr_paper)
• Metadata extraction + enrichment (extract_metadata)
• Section structure extraction (extract_sections)
• Markdown translation (translate_paper)
• Summary report generation (generate_summary)
• Folder rename by metadata convention (rename_paper_folder)
• arXiv search/download tools
• Zotero local read-only library tools (search, notes, collections, attachments, annotations)

Tool Groups

This server currently registers 25 MCP tools:

• Paper processing: ocr_paper, extract_metadata, extract_sections, translate_paper, generate_summary, rename_paper_folder
• Pipeline: process_paper, start_process_paper_job, get_process_paper_job, cancel_process_paper_job, batch_process_papers, get_batch_status, list_jobs, retry_job
• arXiv: search_papers, download_paper
• Zotero read-only: zotero_search_items, zotero_get_item_metadata, zotero_get_item_fulltext, zotero_get_collections, zotero_get_collection_items, zotero_get_tags, zotero_get_recent, zotero_get_annotations, zotero_get_notes

Requirements

• Python >=3.11 (project pin: 3.13.7, see .python-version)
• MinerU token (for OCR)
• OpenAI-compatible LLM API key (default endpoint is DeepSeek)
• Zotero desktop app with local HTTP API enabled (read-only integration)

Configuration Guides:

• How to configure DeepSeek API
• How to configure Zotero local read-only access
• How to configure MinerU OCR
• How to configure MCP in AI Agents (Cursor, Claude, Copilot, Cline)

Installation Modes

There are two ways to install and use this MCP server. Developer mode is recommended, as the installed mode has not been fully tested yet.

Mode 1: Developer Mode (Recommended)

Clone the repository from GitHub, fill in a local .env file, then point your MCP client to the local project directory.

Step 1: Clone the repo and set up the environment:

git clone https://github.com/your-org/fro-wang-academic-tools-mcp.git
cd fro-wang-academic-tools-mcp
uv python pin 3.13.7
uv venv --python 3.13.7
uv sync --extra dev
Copy-Item .env.example .env
# Edit .env and fill in your keys

Edit .env at least for:

• LLM_API_KEY
• MINERU_API_KEY_1
• ZOTERO_LOCAL=true for local read-only Zotero access

Step 2: Configure your MCP client to run the server from the local project path.

The key advantage of this mode is that all credentials are stored in a local .env file — no need to embed them in the MCP client config.

Codex (~/.codex/config.toml)

[mcp_servers.academic_tools]
command = "uv"
args = ["--directory", "/absolute/path/to/mcps/fro-wang-academic-tools-mcp", "run", "fro-wang-academic-tools-mcp"]
startup_timeout_sec = 30.0

Environment variables are read from the .env file in the project directory, so you do not need to set them again in the MCP config.

JSON-based MCP clients (Cursor, Claude Desktop, Cline, etc.)

{
  "mcpServers": {
    "academic_tools": {
      "command": "uv",
      "args": ["--directory", "/absolute/path/to/mcps/fro-wang-academic-tools-mcp", "run", "fro-wang-academic-tools-mcp"]
    }
  }
}

Claude Code

claude mcp add academic-tools -- uv --directory /absolute/path/to/mcps/fro-wang-academic-tools-mcp run fro-wang-academic-tools-mcp

---

Mode 2: Installed Mode via uv tool install (Not Yet Fully Tested)

Warning: This mode has not been fully tested. Use developer mode if you encounter issues.

Install the package directly as a uv tool:

uv tool install fro-wang-academic-tools-mcp

In this mode, the package is installed to a managed location and you cannot easily edit its internal .env. Instead, all credentials must be passed via the MCP client's env block.

Codex (~/.codex/config.toml)

[mcp_servers.academic_tools]
command = "fro-wang-academic-tools-mcp"
args = []
startup_timeout_sec = 30.0

[mcp_servers.academic_tools.env]
LLM_API_KEY = "your_llm_key"
LLM_BASE_URL = "https://api.deepseek.com"
LLM_MODEL = "deepseek-chat"
MINERU_API_KEY_1 = "your_mineru_key"
ZOTERO_LOCAL = "true"

JSON-based MCP clients (Cursor, Claude Desktop, Cline, etc.)

{
  "mcpServers": {
    "academic_tools": {
      "command": "fro-wang-academic-tools-mcp",
      "args": [],
      "env": {
        "LLM_API_KEY": "your_llm_key",
        "LLM_BASE_URL": "https://api.deepseek.com",
        "LLM_MODEL": "deepseek-chat",
        "MINERU_API_KEY_1": "your_mineru_key",
        "ZOTERO_LOCAL": "true"
      }
    }
  }
}

Optionally, you can also point to an external env file instead of embedding keys inline:

ACADEMIC_TOOLS_ENV_FILE=/absolute/path/to/.env

Long-Running Jobs (Recommended for MCP Clients)

Many MCP clients apply a ~60s timeout per tool call. For full pipeline runs, prefer async jobs:

1. Start job with start_process_paper_job (returns immediately with job_id)
2. Poll with get_process_paper_job(job_id)
3. Optional cancel via cancel_process_paper_job(job_id)

This avoids client timeout while OCR/LLM stages continue in the background.

Development

Run the server locally for testing:

uv run fro-wang-academic-tools-mcp

or

uv run python -m academic_tools

Run tests and lint:

uv run pytest
uv run ruff check src

Check Python version used by the project env:

uv run python -V
uv run python -c "import sys; print(sys.executable)"

Common Issues

• uv init says project already initialized:
  • Expected behavior. This repo already has pyproject.toml.
• uv sync --extra dev warns about VIRTUAL_ENV mismatch:
  • You likely activated another project's venv. Deactivate it and run again in this folder.
• Readme file does not exist: README.md during build:
  • This file is required by [project].readme in pyproject.toml. Keep README.md in project root.

Project Entry Points

• CLI script: fro-wang-academic-tools-mcp (defined in pyproject.toml)
• Module entry: python -m academic_tools
• Server wiring: src/academic_tools/server.py