Quick Verdict
Paper Search MCP is a strong infrastructure choice for developers and researchers who want one local interface for discovering scholarly metadata across many sources and retrieving an open-access PDF when one is legitimately available. The openags/paper-search-mcp repository is active and not archived. As of July 21, 2026, PyPI and the Git tag both identify 0.1.4 as the current version. The package requires Python 3.10 or newer and its source code is distributed under the MIT License.
Its central benefit is transparency rather than magical access. search_papers runs selected connectors concurrently and deduplicates records, while download_with_fallback tries source-native delivery, open repositories, and Unpaywall. A metadata match is not proof that a full text is available, and neither the package nor its MIT license grants rights to redistribute papers.
Best For
- Developers building literature-search agents, MCP clients, or reproducible research pipelines.
- Researchers who need broad discovery, DOI recovery, and lawful open-access resolution.
- Command-line users who want JSON search results and plain-text extraction in scripts.
- Teams prepared to monitor upstream API failures, rate limits, and source-specific terms.
- Not ideal for anyone expecting complete coverage of commercial indexes or guaranteed PDFs for every result.
Key Features
- Unified multi-source search: connectors include arXiv, PubMed, bioRxiv, medRxiv, Semantic Scholar, Crossref, OpenAlex, PMC, CORE, Europe PMC, dblp, OpenAIRE, DOAJ, Zenodo, HAL, and others.
- Open-access-first retrieval: the fallback path checks a native source, open repositories, and Unpaywall before any separately enabled restricted connector.
- Normalized records and deduplication: the package maps results to a common
Paperstructure and prioritizes DOI, then title and author, when removing duplicates. - Three usable interfaces: a stdio MCP server, the
paper-searchCLI, and a Claude Code Skill that calls the same underlying package. - Optional credentials: keys can improve rate limits or record access for selected providers. Unpaywall requires an email; other connectors have distinct requirements.
- Auditable Python package: PyPI publishes wheels and source distributions, while the GitHub repository exposes connector code, tests, issues, and an MIT license.
Use Cases
For an initial literature review, run a bounded query against several appropriate sources, preserve source and error metadata, and export a candidate list for manual screening. Verify titles, authors, publication dates, corrections, and retractions before citing anything. For evidence extraction rather than connector-level control, Elicit provides a more finished research interface. For citation-neighborhood exploration, Connected Papers is easier to visualize.
In an agent workflow, place downloads in a dedicated workspace, cap results per provider, validate file types, and avoid automatic citation of PDFs the model has not actually read. Medical searches require especially strict review: this tool does not replace a documented search strategy, professional databases, dual screening, or clinical judgment. Do not expose patient data, unpublished manuscripts, institutional credentials, or sensitive search logs to a shared agent configuration.
Pricing
The repository and Python package are free. Version 0.1.4 was uploaded to PyPI on July 2, 2026, and the repository carries a matching v0.1.4 tag; there is no separate latest GitHub Release object. Installation options include uvx paper-search-mcp, uv tool install paper-search-mcp, pip install paper-search-mcp, Smithery, Docker, and source execution.
MIT licensing covers the software code, not third-party metadata, abstracts, publisher pages, or downloaded PDFs. Upstream providers may impose API quotas, attribution duties, institutional-access rules, or separate charges. Optional keys can improve reliability, but “free package” must not be interpreted as “all scholarly content is free to reuse.”
Pros
- MCP, CLI, and Skill interfaces share one connector implementation.
- Concurrent search, normalized output, and deduplication reduce integration work.
- The capability matrix distinguishes search, download, and read support instead of presenting every source as equivalent.
- Open-access-first retrieval is a better default than assuming a metadata index hosts full text.
- The active MIT-licensed repository is inspectable and extensible.
Cons
- Connector reliability depends on many independent upstream services.
- Google Scholar, SSRN, BASE, CiteSeerX, and other sources can fail because of bot detection, IP policy, redirects, or downtime.
- The default stdio transport is designed for local clients, not a hardened multi-user hosted service.
- Download and text extraction increase exposure to malformed or unexpectedly large files.
- Optional restricted-source support creates legal and institutional-policy risks and should not be enabled by default.
Alternatives
| Tool | Best for | Main difference |
|---|---|---|
| arXiv MCP Server | arXiv-focused search and local paper workflows | Narrower source scope makes behavior easier to predict |
| Elicit | Guided literature screening and evidence tables | Finished research application rather than a self-hosted connector framework |
| Connected Papers | Exploring a seed paper’s citation neighborhood | Visual discovery, not a multi-source download chain |
| Consensus | Natural-language questions over research | Focuses on answers and summaries rather than local MCP control |
| Perplexity | Broad web research and quick leads | Wider web scope but less deterministic academic retrieval |
FAQ
What are the current version and license?
As of July 21, 2026, PyPI and Git tags show 0.1.4. It requires Python 3.10+, and the openags repository identifies the code license as MIT.
Is Paper Search MCP only an MCP server?
No. It exposes the paper-search-mcp stdio server and the paper-search CLI, and the repository documents a Claude Code Skill installation.
Can it work without API keys?
Many public connectors can, but anonymous quotas and stability vary. Unpaywall requires an email, while keys for services such as CORE or Semantic Scholar can improve limits or coverage.
Does every search result include a downloadable PDF?
No. PubMed, Crossref, OpenAlex, and other indexes may provide metadata only. Retrieval depends on source capability, open-access status, and publisher permission.
Does the MIT License let me redistribute downloaded papers?
No. MIT applies to the project code. Authors, publishers, repositories, database terms, and applicable copyright law govern papers and metadata.
Is it sufficient for a medical systematic review?
It can support discovery and initial screening, but it does not replace professional database searches, documented query strategies, human verification, bias assessment, or full-text review.
Bottom Line
Paper Search MCP is an active, useful scholarly-retrieval layer, not a universal paper-download service. Choose it when you need MCP, CLI, and Skill access to the same broad set of connectors. Pin and audit the package, enable only compliant sources, isolate downloads, protect credentials, and keep human source and rights verification in the final research workflow.