Knowra guides

A remote arXiv MCP server you add with one URL

Knowra is a hosted MCP server for arXiv. There is nothing to install or run: add the server URL to your MCP client and your agent can search arXiv by topic and exact announcement date, then read a Scout report on each paper’s methods, evidence and stated limitations.

https://mcp.knowra.app/mcp

Transport: Streamable HTTP. Authentication: OAuth, or a personal API key sent as a bearer token.

Add it to your client

Claude (web or desktop). Find Knowra in the connector directory, or open Settings → Connectors → Add custom connector, paste the URL and choose Connect. Claude guide →

Claude Code.

claude mcp add --transport http --scope user knowra https://mcp.knowra.app/mcp

Then run /mcp in Claude Code, select knowra and choose Authenticate.

ChatGPT (web or desktop, paid plan). Customize → Plugins → Add → Add custom MCP server. Paste the URL, select OAuth, choose Create as a plugin, then Connect Knowra. ChatGPT guide →

Codex CLI.

codex mcp add knowra --url https://mcp.knowra.app/mcp
codex mcp login knowra

Cursor. Add this to ~/.cursor/mcp.json. If the in-app sign-in does not open, use a personal API key from your Knowra account:

{
  "mcpServers": {
    "knowra": {
      "url": "https://mcp.knowra.app/mcp",
      "headers": { "Authorization": "Bearer YOUR_KNOWRA_API_KEY" }
    }
  }
}

Any other remote MCP client. Add the URL with Streamable HTTP. Use OAuth if the client supports it; otherwise send a personal API key in the Authorization header. Keep keys out of URLs, chats and tool arguments.

Step-by-step instructions for every client are in the MCP guide.

What your agent gets

Tool What it does
search_papers Searches arXiv titles, authors and abstracts by topic, with optional from / to dates (UTC, the arXiv announcement date). Returns paper IDs and links.
get_scout Reads a structured report on a paper: problem, method, what was tested, key evidence, verification points and the authors’ stated limitations.
get_usage Shows your Scout allowance and reset time.

All three tools are read-only.

An example

We asked an agent in Cursor: “Use Knowra to find yesterday’s arXiv papers on LLM agent memory and read the Scout for one.”

The agent searched the previous UTC day, listed candidates such as KEEPER, KITE and ASTC, picked Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents (arXiv 2610.41056), and summarized its Scout: the problem (memory managers must accept real updates but reject poisoning), the method (an evidence-conditioned accuracy metric plus one-shot SFT), the results, and the limits the authors state, including a poisoning test set of 47 attacked samples. We ran the same request with four different models; every one reported the same facts from the Scout.

Prefer to run it yourself?

Open-source arXiv MCP servers on GitHub run locally and fetch from the arXiv API. They are a good fit if you want to self-host or read full PDFs locally. Knowra is the hosted option: one URL, exact announcement-date search across every arXiv subject, and Scout reports your agent can read before it recommends a paper.

FAQ

What is the Knowra MCP server URL? https://mcp.knowra.app/mcp, over Streamable HTTP.

Which arXiv papers are covered? Every arXiv subject, from August 1, 2025 onward, updated with each daily announcement.

Can I search by date? Yes. search_papers takes from and to in YYYY-MM-DD (UTC), both inclusive, up to one calendar month per request. The date is the arXiv announcement date. See how to search arXiv by date.

Does Scout replace reading the paper? No. Scout helps your agent decide which papers deserve a close read and what to check in them. Verify specific claims, numbers and figures in the original.

How do I disconnect? Revoke the client under Connected apps in your Knowra account, or revoke the API key.

See what Knowra does →