mcp-scholarly MCP server

mcp-scholarly MCP server

A MCP server to search for accurate academic articles. More scholarly vendors will be added soon.

Search tools

• search-arxiv — arXiv search (no key needed)
• search-google-scholar — Google Scholar via the scholarly library (free proxy pool)
• search-google-web — Google web search via the SerpBase API. Optional; only registered when SERPBASE_API_KEY is set. Get a key at https://serpbase.dev/dashboard/api-keys (free tier available).

Components

Tools

The server implements one tool:

• search-arxiv: Search arxiv for articles related to the given keyword.
  • Takes "keyword" as required string arguments

Quickstart

Install

Claude Desktop

On MacOS: ~/Library/Application\ Support/Claude/claude_desktop_config.json On Windows: %APPDATA%/Claude/claude_desktop_config.json

Development/Unpublished Servers Configuration

``` "mcpServers": { "mcp-scholarly": { "command": "uv", "args": [ "--directory", "/Users/adityakarnam/PycharmProjects/mcp-scholarly/mcp-scholarly", "run", "mcp-scholarly" ] } } ```

Published Servers Configuration

``` "mcpServers": { "mcp-scholarly": { "command": "uvx", "args": [ "mcp-scholarly" ] } } ```

or if you are using Docker

Published Docker Servers Configuration

``` "mcpServers": { "mcp-scholarly": { "command": "docker", "args": [ "run", "--rm", "-i", "mcp/scholarly" ] } } ```

Installing via Smithery

To install mcp-scholarly for Claude Desktop automatically via Smithery:

npx -y @smithery/cli install mcp-scholarly --client claude

Development

Building and Publishing

To prepare the package for distribution:

1. Sync dependencies and update lockfile:

uv sync

2. Build package distributions:

uv build

This will create source and wheel distributions in the dist/ directory.

3. Publish to PyPI:

uv publish

Note: You'll need to set PyPI credentials via environment variables or command flags:

• Token: --token or UV_PUBLISH_TOKEN
• Or username/password: --username/UV_PUBLISH_USERNAME and --password/UV_PUBLISH_PASSWORD

Debugging

Since MCP servers run over stdio, debugging can be challenging. For the best debugging experience, we strongly recommend using the MCP Inspector.

You can launch the MCP Inspector via npm with this command:

npx @modelcontextprotocol/inspector uv --directory /Users/adityakarnam/PycharmProjects/mcp-scholarly/mcp-scholarly run mcp-scholarly

Upon launching, the Inspector will display a URL that you can access in your browser to begin debugging.

Using with zorp

zorp needs a search-capable MCP tool before validate will run. This server satisfies that check, because zorp matches on a search verb in the tool name and these tools are called search-arxiv and search-google-scholar.

zorp-agent --yes \
  --mcp "stdio:scholarly:uv:run:mcp-scholarly" \
  validate "<your research question>"

Or configure it once, so every run picks it up:

# .zorp/mcp.toml
[[server]]
name = "scholarly"
transport = "stdio"
command = "uv"
args = ["run", "mcp-scholarly"]
trust = "sandbox"
timeout_secs = 60

Notes measured against zorp's transport, not assumed:

• search-arxiv answers in about 1 second. zorp's default stdio read budget is 30 seconds, so the default is comfortable. timeout_secs = 60 above is headroom for search-google-scholar, which goes through scholarly and a free proxy pool and is far less predictable.
• Logging goes to stderr. Nothing but JSON-RPC reaches stdout, which is what zorp's newline-delimited framing requires.
• An empty keyword comes back as an MCP tool error rather than an empty result set. zorp cares about that distinction: a failed search that looks like "no prior work" would put a wrong novelty score into an evidence record.
• arxiv returns best-effort matches for any query, including nonsense, so a non-empty result set is not by itself evidence that prior work exists. The tool description says so, since that is the text the model reads.