arxivsub-mcp
# arxivsub-mcp
An [MCP](https://modelcontextprotocol.io) server for **[arXivSub](https://arxivsub.comfyai.app)** — search arXiv and the top AI / CV / ML / robotics / NLP conferences, follow any paper to its related work, build research analytics, and manage a personal library, straight from your AI agent (Claude Desktop, Cursor, Cline, Claude Code, Windsurf, …).
arXivSub summarizes every paper into a structured 6-part breakdown (what it's about, innovations, techniques, datasets, results, limitations) and indexes it for semantic search across **arXiv + CVPR, ICCV, ECCV, ICLR, ICML, NeurIPS, AAAI, MICCAI, CoRL, RSS, ACL, EMNLP, ICRA, IROS, IJCAI, SIGGRAPH, SIGGRAPH Asia**. This server puts that engine directly in your agent's hands.
## Get a free key (7-day trial, no credit card)
1. Sign in at **https://arxivsub.comfyai.app/skills** with email or Google.
2. Copy your API key from the Skills page.
3. Put it in the server config below as `ARXIVSUB_SKILL_KEY`.
New accounts get a **7-day free trial** of the full skill. After that it's part of arXivSub Pro.
## Install
The server runs via `npx` — no manual install needed. Add it to your MCP client's config.
**Claude Desktop** (`claude_desktop_config.json`), **Cursor** (`~/.cursor/mcp.json`), **Windsurf**, **Cline**, etc.:
```json
{
"mcpServers": {
"arxivsub": {
"command": "npx",
"args": ["-y", "arxivsub-mcp"],
"env": { "ARXIVSUB_SKILL_KEY": "your_key_here" }
}
}
}
```
**Claude Code:**
```bash
claude mcp add arxivsub --env ARXIVSUB_SKILL_KEY=your_key_here -- npx -y arxivsub-mcp
```
**Codex CLI:**
```bash
codex mcp add arxivsub --env ARXIVSUB_SKILL_KEY=your_key_here -- npx -y arxivsub-mcp
```
Or add it to `~/.codex/config.toml` directly:
```toml
[mcp_servers.arxivsub]
command = "npx"
args = ["-y", "arxivsub-mcp"]
env = { ARXIVSUB_SKILL_KEY = "your_key_here" }
```
Restart your client and the tools below become available.
## claude.ai (remote connector)
On [claude.ai](https://claude.ai) there's nothing to install, add a hosted connector:
1. **Settings → Connectors → Add custom connector**
2. **Name**: `arxivsub`
3. **Remote MCP server URL** (replace `YOUR_KEY` with your key from https://arxivsub.comfyai.app/skills):
```
https://qtevnmgyobilaanrzidq.supabase.co/functions/v1/mcp?key=YOUR_KEY
```
4. Leave the OAuth fields empty and click **Add**.
This is a hosted [Streamable HTTP](https://modelcontextprotocol.io) MCP endpoint, a thin adapter over the arXivSub gateway exposing the same tools. Source: [`supabase/functions/mcp/index.ts`](supabase/functions/mcp/index.ts).
## Tools
| Tool | What it does |
|------|--------------|
| `search_papers` | Semantic + keyword search over arXiv and conferences. Filters: venues, recency, year, `has_code`. Returns each paper's 6-part summary, authors, PDF, and code link. |
| `find_similar_papers` | Nearest-neighbour "related work" for a paper you already found. Chainable for a mini literature review. |
| `research_insight` | Pivot analytics (the website's Custom Chart engine): trends, rankings, breakdowns. *Pro / trial.* |
| `library_save` | Save/update a paper in your personal library — notes, reading status, collections. |
| `library_list` | List your saved papers and collections. |
Just ask your agent naturally — *"find recent papers on LLM safety with code"*, *"more work like this one"*, *"how has interest in diffusion models trended by month"*, *"save this to my "to read" collection"* — and it will pick the right tool.
## Notes
- **Case-sensitive venues:** `arxiv`, `CVPR`, `ICCV`, `ECCV`, `ICLR`, `ICML`, `NeurIPS`, `AAAI`, `MICCAI`, `CoRL`, `RSS`, `ACL`, `EMNLP`, `ICRA`, `IROS`, `IJCAI`, `SIGGRAPH`, `SIGGRAPH Asia`.
- **Quota:** each search / similar / insight call uses one daily quota unit; library calls are free. The remaining quota is returned with every result.
- **Privacy / security:** the server only holds your skill key and talks to the arXivSub gateway over HTTPS. It has no direct database access and no service credentials.
## Development
```bash
npm install
npm run build # tsc -> dist/
ARXIVSUB_SKILL_KEY=... node dist/index.js # run over stdio
```
## License
MIT
TDQS
Scored across 5 tools
Each tool targets a distinct function: search_papers for query-based discovery, find_similar_papers for similarity expansion, research_insight for aggregate analytics, and library_save/library_list for personal library management. There is no ambiguity between them.
Most tools follow a verb_noun pattern (search_papers, find_similar_papers), but library_save and library_list invert this to noun_verb, and research_insight is ambiguous. The underscore-separated lowercase style is consistent, but the word order is not uniform.
At 5 tools, the set is well-scoped for the server's purpose: search, similar, analytics, and library operations. Each tool earns its place without redundancy or bloat.
The surface covers core workflows (search, similar, trends, save, list), but library_delete is explicitly absent, forcing removals to be handled externally. A direct get-by-ID tool is also missing, though search can compensate. Minor gaps that agents can work around.