🧠 arxiv-daily-mcp

A Model Context Protocol (MCP) server that wraps an arXiv CLI workflow. Search arXiv in 🇰🇷 Korean, 🇨🇳 Chinese, 🇯🇵 Japanese, or 🇺🇸 English — and save findings to a local Markdown vault.

🌍 Read in your language: English · 한국어 · 中文 · 日本語 · Español · Français · हिन्दी · العربية

🚀 Quickstart · 🛠 Tools · 🎬 Demo · 🆚 vs other arXiv MCPs · 🏗 Architecture · 🧪 Development · 🌐 Multilingual

✨ What this is

A simple MCP server that exposes a useful arXiv workflow as LLM-callable tools. The core idea:

Take a CLI script that you already run every day, and wrap it in MCP so you can invoke it with natural language from any MCP client.

If you've ever found yourself typing the same curl / python / arxiv query into a terminal every morning — this is what that looks like as an MCP tool.

Related MCP server: arxiv-search MCP Server

🛠 Tools (4)

get_today_papers(category, max_results=5)

• category: "ai" (cs.AI+cs.CL+cs.LG) · "business" (econ.GN+q-fin) · "all" (전체)

• Returns: Markdown card-news list of today's submissions

search_papers(keyword, max_results=5)

• keyword: any text in 🇰🇷 Korean, 🇨🇳 Chinese, 🇯🇵 Japanese, or 🇺🇸 English

• Auto-translates non-English keywords to English for arXiv search

• Returns: Markdown card-news of results, with translation note

save_paper_note(arxiv_id, note="")

• arxiv_id: "2507.01234" or full URL https://arxiv.org/abs/2507.01234

• note: your commentary (optional)

• Returns: path to saved file (default: ~/memos/YYYY-MM-DD_arxiv_<id>.md)

list_my_notes(keyword="", limit=10)

• keyword: filter by title / content (optional)

• Returns: list of recently saved note files

🚀 Quickstart

Install

pip install arxiv-daily-mcp

Or from source:

git clone https://github.com/nerin81-netizen/arxiv-daily-mcp
cd arxiv-daily-mcp
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Register with your MCP client

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "arxiv-daily": {
      "command": "arxiv-daily-mcp"
    }
  }
}

Cursor / Cline

.cursor/mcp.json or workspace settings:

{
  "mcpServers": {
    "arxiv-daily": {
      "command": "arxiv-daily-mcp"
    }
  }
}

Any MCP-compatible client

This server uses the standard stdio transport and exposes 4 tools via JSON-RPC — it works with any client that follows the MCP spec.

🎬 Demo

Korean keyword (auto-translated to English)

👤 에이전트 관련 논문 찾아줘

🤖 Claude:
🔍 '에이전트' 검색 결과 — 🇰🇷 한국어 → 영어: 'agent' (5편)

1. **Distributed Attacks in Persistent-State AI Control**
   👥 Josh Hills, Ida Caspary, Asa Cooper Stickland
   📅 2026-07-02 | 🏷️ cs.AI
   🔗 https://arxiv.org/abs/2607.02514v1

Chinese keyword

👤 智能体 相关论文

🤖 Claude:
🔍 '智能体' 검색 결과 — 🇨🇳 中文 → 영어: 'agent' (5편)

Japanese keyword

👤 エージェント 論文を探して

🤖 Claude:
🔍 'エージェント' 검색 결과 — 🇯🇵 日本語 → 영어: 'agent' (5편)

Save + retrieve

👤 첫 번째 논문 메모해줘 — 멀티에이전트 보안 흥미롭다

🤖 Claude:
✅ 저장 완료: `~/memos/2026-07-06_arxiv_2607.02514.md`

👤 내 메모 보여줘

🤖 Claude:
📚 내 arXiv 노트 (1개)
  · `2026-07-06_arxiv_2607.02514.md` — Distributed Attacks in Persistent-State AI Control

🆚 How it differs from existing arXiv MCPs

GitHub has 466+ arXiv MCP repositories (as of 2026-07-05). The most popular, blazickjp/arxiv-mcp-server, is excellent — but built for a different audience.

Three reasons it can coexist

1. Different language audience. The 2,933-star leader is English-only. This server is designed for non-English researchers first.

2. Different output shape. The leader returns raw JSON / data tables. This server returns card-news Markdown — readable directly inside any chat UI.

3. Different workflow. The leader is search-and-read. This server includes archival: save a paper to your local Markdown vault with one sentence, then list_my_notes to recall it later.

🌐 Multilingual keywords

The translation dictionary ships with the package — no API call, no latency, no rate limit.

Detection algorithm:

1. If the input contains Hangul (가-힣) → Korean lookup

2. Else if it contains Kana (ぁ-ヿ) → Japanese lookup

3. Else if it contains CJK ideographs (一-鿿) → Chinese lookup

4. Else → English (passed through unchanged)

Unknown keywords are passed through as-is — so a partial match still searches arXiv directly.

🏗 Architecture

┌─────────────────┐      stdio/JSON-RPC      ┌──────────────────────┐
│  LLM Client     │ ◀──────────────────────▶ │  arxiv-daily-mcp     │
│  (Claude/Cursor) │                          │  (FastMCP server)    │
└─────────────────┘                          └──────────┬───────────┘
                                                        │ urllib
                                                        ▼
                                                 ┌──────────────┐
                                                 │  arXiv API   │
                                                 │  export.arxiv│
                                                 └──────────────┘
                                                        │
                                                        ▼ (save_paper_note)
                                                 ┌──────────────┐
                                                 │  ./memos/    │
                                                 │  *.md vault  │
                                                 └──────────────┘

Key design choices:

• stdio transport — MCP standard

• FastMCP — Anthropic's official Python SDK

• urllib only — zero external HTTP dependencies (works in any minimal Python environment)

• In-process translation dictionary — no LLM call, no rate limit, no per-keyword latency

• Configurable MEMO_DIR — set via env var for Docker / multi-user setups

🧪 Development

# Install dev dependencies
pip install -e ".[dev]"

# Run MCP Inspector (browser UI)
npx @modelcontextprotocol/inspector arxiv-daily-mcp

# Test stdio JSON-RPC directly
echo '{"jsonrpc":"2.0","method":"tools/list","id":1}' | arxiv-daily-mcp

Unit test the translation layer

from arxiv_daily_mcp.server import _translate_keyword

assert _translate_keyword("에이전트") == ("agent", "ko")
assert _translate_keyword("智能体") == ("agent", "zh")
assert _translate_keyword("エージェント") == ("agent", "ja")
assert _translate_keyword("agent") == ("agent", "en")

🌍 Roadmap

• v0.1.0 — 4 tools, stdio, multilingual (75 keywords)

• v0.2.0 — weekly_summary tool (Friday digest)

• v0.3.0 — More categories (cs.RO, cs.CV, q-fin.ST)

• v0.4.0 — PyPI release

• v1.0.0 — HTTP transport (remote access)

• v1.1.0 — More languages (Spanish, French, German, Hindi)

• v2.0.0 — Note format plugins (Notion, Apple Notes, Bear)

🤝 Contributing

Issues and PRs welcome. Particularly useful:

• 🌏 More language keywords (especially in cs.RO, q-fin, stat.ML)

• 📂 New paper categories

• 📝 Note format options (Notion API, Apple Notes, Bear)

• 🐛 Bug reports with reproduction steps

📜 License

MIT

🙏 Credits

• Built with Anthropic MCP Python SDK

• Paper data from arXiv.org (open API)

• Inspired by the modelcontextprotocol/servers reference collection

🌟 If this MCP saved you a few minutes of research workflow, consider starring the repo — it helps others find it.