# Arxiv [Health: Active]

**Category:** 🔬 Research  
**Repository:** https://github.com/pipeworx-io/mcp-arxiv  
**GitHub Stars:** 0  
**npm Downloads (last month):** 439  
**Views:** 7  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/arxiv-2

## Description
arXiv MCP — preprint server search (free, no auth)

## Tools
Capabilities this server exposes over MCP:

- **search_papers** — Search arXiv preprints. Plain text searches all fields; use prefixes for targeted queries: au:hinton (author), ti:transformer (title), abs:diffusion (abstract), cat:cs.AI (category), all:quantum (any field). Combine with AND/OR/ANDNOT, e.g., "ti:transformer AND cat:cs.LG". Returns id, title, authors, abstract, categories, published date, PDF URL.
- **get_paper** — Fetch a single arXiv paper by its ID (e.g., "2310.06825", "2310.06825v2", or legacy "cs.CL/0301001"). Returns full metadata: title, authors, abstract, categories, DOI (if linked), PDF URL.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "arxiv": {
    "command": "npx",
    "args": ["-y","@pipeworx/mcp-arxiv"]
  }
}
```

## Documentation

## What Arxiv does

Arxiv provides MCP access to arXiv preprints for literature discovery and paper lookup. The server exposes two focused tools: `search_papers` searches across arXiv records, while `get_paper` retrieves metadata for one paper. Search results include identifiers, titles, authors, abstracts, categories, publication dates, and PDF URLs. Individual lookups can also include a linked DOI.

The integration is unofficial and connects to arXiv through Pipeworx. It does not require an arXiv API key or authentication for the hosted service's initial calls.

## How it works

The hosted MCP endpoint is `https://gateway.pipeworx.io/arxiv/mcp`. An MCP client can connect to that URL and discover the Arxiv tools. The underlying data source is arXiv's public Atom API at `https://export.arxiv.org/api/query`.

A local npm package, `@pipeworx/mcp-arxiv`, acts as a stdio bridge. Your MCP client starts it with `npx`, and the bridge forwards requests to the same hosted gateway, so it requires network access and has no offline mode. The gateway also offers direct HTTP tool calls, including a POST endpoint for `arxiv_search_papers`.

## Setup and configuration

For a hosted connection, add the following MCP server entry:

```json
{
  "mcpServers": {
    "arxiv": {
      "url": "https://gateway.pipeworx.io/arxiv/mcp"
    }
  }
}
```

For a local stdio connection, use:

```json
{
  "mcpServers": {
    "arxiv": {
      "command": "npx",
      "args": ["-y", "@pipeworx/mcp-arxiv"]
    }
  }
}
```

Anonymous npm-bridge usage is limited to 50 calls per day. Setting `PIPEWORX_API_KEY` with a key from Pipeworx raises the limit to 200 calls per day. Set `PIPEWORX_META_TOOLS=1` if the local bridge should also list Pipeworx's shared gateway tools. Requests behind an HTTPS proxy may require `NODE_USE_ENV_PROXY=1` with a compatible Node.js release.

## Tools and capabilities

- `search_papers(query, max_results?, sort_by?, sort_order?)` searches arXiv records. Plain text searches all fields; prefixes target specific fields:
  - `au:` author
  - `ti:` title
  - `abs:` abstract
  - `cat:` category
  - `all:` any field
- Boolean operators `AND`, `OR`, and `ANDNOT` can combine search conditions, such as `ti:transformer AND cat:cs.LG`.
- `get_paper(arxiv_id)` accepts modern IDs such as `2310.06825` and versioned IDs such as `2310.06825v2`, along with legacy IDs such as `cs.CL/0301001`.
- Direct HTTP access is available through the Pipeworx tool endpoint when an MCP client is not being used.

## Limitations and notes

The npm bridge depends on the hosted Pipeworx gateway and cannot operate offline. A pack endpoint also lists shared Pipeworx meta-tools alongside the Arxiv tools; setting `PIPEWORX_META_TOOLS=1` enables those tools for the local bridge. The integration is independent and is not affiliated with, endorsed by, or published by arXiv.

The project is distributed under the MIT license.

_Full upstream README: https://allmcps.com/mcp/arxiv-2/readme_