hf-papers-mcp

An MCP server that lets your assistant read Hugging Face Papers — search them, pull the daily curated list, fetch metadata, and generate citations.

Ask "what shipped on (your interest) this week?" and get real papers with upvote counts and abstracts, instead of whatever was in the model's training data.

(Recorded with vhs via vhs demo.tape — the session runs bin/demo.js, a real MCP client against the live API.)

Install

Point your MCP client (Claude Desktop, Claude Code, or any MCP host) at it:

{
  "mcpServers": {
    "hf-papers": {
      "command": "npx",
      "args": ["-y", "hf-papers-mcp"]
    }
  }
}

Or from source: git clone https://github.com/QEbellavita/hf-papers-mcp and point command: node, args: [/absolute/path/to/hf-papers-mcp/server.js] at the checkout.

Pure Node (>= 18), one dependency (@modelcontextprotocol/sdk), no Python. No API key needed — everything is a public read. If HF_TOKEN is set (or you're logged in via hf auth login), it's sent along, which helps with rate limits.

Related MCP server: semantic-scholar-mcp

Tools

Example

> what were the top papers on Hugging Face yesterday?

  [1] Domino: Decoupling Causal Modeling from Autoregressive Drafting
      in Speculative Decoding
      arXiv: 2605.29707  |  Upvotes: 152

Bulk digests

Pull a date range in one go:

node bin/daily.js 2026-07-01 2026-07-14

Writes JSON keyed by date to /tmp/, and records the window in .hf_papers_last_run.json so a bare node bin/daily.js picks up where the last run finished.

Tests

npm test                      # hermetic — no network
HF_PAPERS_E2E=1 npm test      # adds live API tests

The default suite is offline and covers tool definitions plus a real MCP handshake over stdio. The live tests hit the Hugging Face API and assert that responses come back as parsed JSON — a regression guard, because diagnostics written to stdout used to corrupt the JSON stream and silently degrade structured responses into raw strings.

Notes

Every call is a direct HTTPS request to the public Hugging Face / arXiv APIs — no daemon, no cached state, nothing persisted beyond the optional bin/daily.js run marker. Diagnostics go to stderr, since stdout is the MCP transport and anything else written there breaks the protocol.

Earlier versions shelled out to a Python sidecar (scripts/paper_manager.py via uv); v1.1.0 ported the six exposed operations to native Node, so cold-start latency dropped from ~12s to nothing and the uv/Python requirement is gone.

Citing

If this server is useful in your research workflow, cite it via the repo's CITATION.cff (GitHub's "Cite this repository" button), or the Zenodo DOI, 10.5281/zenodo.21484748 — a concept DOI that always resolves to the latest archived version.

Licence

MIT — see LICENSE.