Scholar Feed

Semantic and citation search CS/AI/ML research papers.

Search 600,000+ CS/AI research papers with citation graphs, full text, embeddings, and BibTeX. Papers are scraped from ArXiv daily and ranked by novelty and quality with AI.

MCP connector

Server URL: https://mcp.scholarfeed.org/mcp
Login: OAuth
Protocol: MCP 2025-11-25
Instructions: Scholar Feed is a research copilot over 600k+ CS/AI/ML papers, not just a search index. A single call returns roughly what a web search would; the differentiated value is the citation graph and the rising-work signal layered on top. For any non-trivial research request, do not stop at the first search.
Deep-research loop:
(q=...) to find anchor papers for the topic.
(anchor_paper_id=<anchor>) to surface the canonical prior art that semantic search misses.
(arxiv_id=<anchor>, direction="cited_by") to find newer work that builds on it. This is how you reach recent papers a model cannot recall from training.
(q=..., sort="trending") or days=<N> for the rising frontier. Keep q on every search: sort= reranks the matches for a topic, it is not a topic-free feed.
(arxiv_ids=["id1","id2","id3"], sections=["method","results"]) to read your top few hits, not just one, before answering. Read a shortlist in ONE call (up to 8 papers) and ask only for the sections you need: sections=["all"] is about 13.5KB per paper, so eight of those is about 108KB of context. Use (arxiv_id="id") for a single paper.
From what you read, look up the baselines and leaderboards those papers name. The paper everyone benchmarks against is often modestly cited and ranked below the newest work, so chase named baselines rather than only taking the freshest result.

Tools

Reads data17

Changes data7

Deletes or can't be undone3

Skills

None in the store listings or on the server

MCP connector

Server URL

https://mcp.scholarfeed.org/mcp

OAuth

Protocol

MCP 2025-11-25

Instructions

Scholar Feed is a research copilot over 600k+ CS/AI/ML papers, not just a search index. A single call returns roughly what a web search would; the differentiated value is the citation graph and the rising-work signal layered on top. For any non-trivial research request, do not stop at the first search.

Deep-research loop:

(q=...) to find anchor papers for the topic.

(anchor_paper_id=<anchor>) to surface the canonical prior art that semantic search misses.

(arxiv_id=<anchor>, direction="cited_by") to find newer work that builds on it. This is how you reach recent papers a model cannot recall from training.

(q=..., sort="trending") or days=<N> for the rising frontier. Keep q on every search: sort= reranks the matches for a topic, it is not a topic-free feed.

(arxiv_ids=["id1","id2","id3"], sections=["method","results"]) to read your top few hits, not just one, before answering. Read a shortlist in ONE call (up to 8 papers) and ask only for the sections you need: sections=["all"] is about 13.5KB per paper, so eight of those is about 108KB of context. Use (arxiv_id="id") for a single paper.

From what you read, look up the baselines and leaderboards those papers name. The paper everyone benchmarks against is often modestly cited and ranked below the newest work, so chase named baselines rather than only taking the freshest result.

Tools