Scholar MCP

Find the paper. Follow the evidence. Build the field.
Documentation · English · 简体中文

Scholar MCP gives your AI agent a research library. It searches scholarly sources in parallel, with optional keys for additional indexes, follows citations and related work, reads the original paper down to its figures and tables, and keeps what matters in a local library with your notes. It works in Claude, Codex, Cursor, VS Code, Gemini CLI, OpenCode and any other MCP client, and from the command line or Python.

📰 News

• November 2026, 0.9.0. scholar-mcp install sets Scholar up in 32 agents, one key file serves all of them, every source searches in its own lane, and scholar-mcp doctor --network checks every link.
• October 2026, 0.8.5. Google Scholar sessions recover on their own, and arXiv falls back to its own search when its API is down.

🚀 Get started

1. Install uv if you do not have it, then install Scholar:

   uv tool install --python 3.12 --find-links https://github.com/Liyux3/scholar-mcp/releases/latest/download/wheel-index.html "scholar-mcp[rerank]"
   

2. Connect it to your agents. Scholar finds the clients on your machine and registers itself, with a backup of every file it touches:

   scholar-mcp install --all
   

   In a client the installer does not know, ask your agent to install Scholar by following llms-install.md. Name clients to choose (scholar-mcp install claude-code codex), add --dry-run to preview the change, or --profile research for the citation graph and paper library tools. scholar-mcp update upgrades the command every connected client runs, and uvx scholar-mcp@latest install CLIENT refreshes a client pinned to a uvx version.

3. Restart the client and ask a research question, for example find recent work on evaluating long-term memory in LLM agents, read the most cited one and compare its benchmarks.

{
  "mcpServers": {
    "scholar": {
      "command": "uvx",
      "args": [
        "--python", "3.12",
        "--find-links", "https://github.com/Liyux3/scholar-mcp/releases/latest/download/wheel-index.html",
        "--from", "scholar-mcp[rerank]", "scholar-mcp"
      ]
    }
  }
}

codex plugin marketplace add Liyux3/scholar-mcp && codex plugin add scholar-mcp@scholar-mcp
claude plugin marketplace add Liyux3/scholar-mcp && claude plugin install scholar-mcp@scholar-mcp

🧰 What your agent can do

Every identifier Scholar returns, a paper URL, or an exact title can be passed back to any tool. Responses carry readable YAML for the model and structured data for the client; scholar://status reports the sources and reranker in use.

🧭 How it works

Semantic sources receive the whole question and keyword sources a compact query tuned per source. Each source answers in its own lane at the pace it allows, and a source that misses the first ranking still joins the final one. Records from OpenAlex, Semantic Scholar, arXiv, Crossref, OpenReview, PubMed, Europe PMC, DBLP, zbMATH, CORE, OpenAIRE, Google Scholar and others merge on shared DOIs, arXiv ids and titles, so each paper appears once with complementary metadata. A reranker (Qwen3 Rerank in the cloud, or a local model) orders the pool, the strongest matches pull in their citations and references, and the combined pool is ranked once more. The architecture page walks through each stage and the backend film follows one question through it in two and a half minutes.

Runtime architecture

Agents call typed MCP tools over stdio or Streamable HTTP. Full text comes from open copies first and, for licensed papers, through your library's proxy. A persistent SQLite library with FTS5 search holds papers, notes and PDF attachments under ~/.scholar-mcp/, with Markdown projections for Obsidian and connectors for Zotero and Notion.

📊 Measured retrieval quality

On the same 50 LitSearch queries, Scholar placed a correct paper in its top five for 62% of queries, against 52% for Exa research-paper search and 46% for BM25.

📚 Your paper library

Saved papers, PDFs and notes live in one local SQLite library and stay searchable across sessions. Obsidian receives a live Markdown projection, Zotero a synchronized collection and Notion a reading list. Notes you edit in Obsidian or Notion flow back, so your agent builds on your changes.

scholar-mcp library export obsidian --collection memory --path ~/Vault
scholar-mcp library sync zotero --collection memory --apply

💻 Command line and Python

The tools work without an agent too. Output is YAML for reading or JSON with --json.

scholar-mcp search "long-term memory benchmarks for LLM agents" --year 2024-2026
scholar-mcp read arXiv:2504.19413 --pages 1-4
scholar-mcp read arXiv:2504.19413 --visual "Figure 1" --image figure1.png
scholar-mcp doctor --network

from scholar_mcp import api

hits = api.search("long-term memory benchmarks for LLM agents", year="2024-2026")
paper = api.info(hits["results"][0]["id"], include="detail,citations")

🔑 Configuration

Scholar works without any keys. Optional keys add coverage and throughput. scholar-mcp keys set NAME saves one in ~/.scholar-mcp/keys.env, which every client reads; a client's env block or your shell also works and takes precedence.

export LIBRARY_PROXY_PREFIX="https://proxy.library.example.edu/login?url="
export LIBRARY_PROXY_COOKIES="$HOME/private/library-cookies.txt"

🔧 Development

git clone https://github.com/Liyux3/scholar-mcp.git
cd scholar-mcp
uv sync --extra dev
uv run pytest            # unit tests; live API tests run with -m integration

A new source is one module in scholar_mcp/sources/ plus one registry entry, and needs no new tool. scholar-mcp doctor --network probes every source, the ranking model and the PDF chain, and runs daily in CI. See ARCHITECTURE.md for how the parts fit and CONTRIBUTING.md for changes, and report security issues through SECURITY.md.

📝 Citation

If Scholar MCP supports your research, please cite it.

@software{li_scholar_mcp_2026,
  author  = {Li, Yuxian},
  title   = {Scholar MCP},
  year    = {2026},
  version = {0.9.0},
  url     = {https://github.com/Liyux3/scholar-mcp},
  license = {Apache-2.0}
}

📄 License

Apache License 2.0. Reuse is welcome; keep the NOTICE file and credit Scholar MCP in what you build, whether a person or an AI agent did the copying.

Bundled CPU reader weights are provided unchanged by prithivida/FlashRank under CC-BY-SA-4.0, with their own attribution notice.