arXiv Research MCP

A read-only, deployable MCP server for scientific literature research.

ChatGPT
   ↓  Streamable HTTP
arXiv Research MCP
   ↓  official Atom API
arXiv scientific papers

The server retrieves and normalizes metadata. ChatGPT remains responsible for explaining physics, comparing papers, and working through equations.

Tools

Every tool is marked read-only and returns typed structured data containing the arXiv ID, normalized title and abstract, authors, categories, dates, DOI, journal reference, abstract URL, and PDF URL when available.

Local setup

Requires Python 3.11 or newer.

py -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"
Copy-Item .env.example .env

Edit ARXIV_USER_AGENT in .env and replace the example contact with your email or project URL. arXiv does not require an API key.

Run locally over stdio:

arxiv-mcp

Run the remote-compatible transport locally:

arxiv-mcp --transport streamable-http --host 127.0.0.1 --port 8000

The endpoints are:

http://127.0.0.1:8000/mcp
http://127.0.0.1:8000/health

For a manual tool test, start the HTTP server, run the official MCP Inspector, and connect it to http://127.0.0.1:8000/mcp:

npx @modelcontextprotocol/inspector@latest

Run tests

pytest
ruff check .

Tests use mocked HTTP responses and do not consume the arXiv API. GitHub Actions runs the same checks on every push and pull request.

Reliability and security

• Fixed allowlisted upstream: https://export.arxiv.org/api/query

• Three-second global request spacing by default

• Bounded retries for timeouts, network failures, HTTP 429, and temporary 5xx responses

• Strict result, offset, category, sort, and arXiv-ID validation

• Safe XML parsing with external entities disabled

• No shell execution, arbitrary URL fetching, secrets, write operations, or user-data storage

• Stateless Streamable HTTP deployment

Deploy on Render

The repository includes a non-root Dockerfile and render.yaml Blueprint.

1. Push this repository to GitHub.

2. Sign in to Render and select New → Blueprint.

3. Connect the GitHub repository and apply render.yaml.

4. In the Render service environment, change ARXIV_USER_AGENT to include your real contact email or repository URL.

5. Wait for /health to pass, then copy the service URL.

The final MCP URL is:

https://<your-render-service>.onrender.com/mcp

The Blueprint uses Render's free instance so deployment does not silently create a billing commitment. Render documents that free services sleep after 15 minutes of inactivity and can take about a minute to restart. That cold start can exceed an MCP client's timeout; upgrade the service to starter or another paid instance before relying on it as an always-available production connector. See the Render free-tier limitations and Blueprint reference.

Connect to ChatGPT

These steps follow the current official OpenAI documentation:

1. Deploy the server and confirm the public HTTPS /mcp URL works in MCP Inspector.

2. In ChatGPT, open Settings → Security and login and enable Developer mode.

3. Open ChatGPT Plugins and select the plus button.

4. Enter:

   • Name: arXiv Research

   • Description: Read-only search and metadata retrieval from the official arXiv API.

   • Connection: public MCP endpoint

   • Server URL: https://<your-render-service>.onrender.com/mcp

   • Authentication: none

5. Create the connection and confirm that exactly five tools are discovered.

6. Start a new chat and add arXiv Research from the tools menu.

OpenAI currently requires a public HTTPS endpoint (or its Secure MCP Tunnel) and Streamable HTTP, typically at /mcp. Developer-mode availability depends on the account and workspace policy. OpenAI's current page does not promise it for every named ChatGPT plan, so the presence of the Developer mode toggle on your account is the authoritative check. If a Plus account does not show it, the server remains usable through MCP Inspector, another MCP client, or the OpenAI API Playground. See OpenAI's current connection guide and MCP server guide.

Example ChatGPT prompts

• Search arXiv for recent quantum gravity papers.

• Find papers about the Bondi-Sachs formalism.

• Search hep-th for path integral approaches to gravity.

• Find papers by Roger Penrose related to general relativity.

• Get the metadata and PDF link for arXiv:2401.12345.

Project structure

.
├── src/arxiv_mcp/
│   ├── arxiv_client.py
│   ├── config.py
│   ├── models.py
│   └── server.py
├── tests/
├── .github/workflows/ci.yml
├── .env.example
├── Dockerfile
├── render.yaml
├── pyproject.toml
└── README.md

MIT licensed.