fulltext-article-downloader

fulltext-article-downloader is a Python package for programmatically downloading the full text of research articles from a DOI, arXiv id, PubMed Central id or OpenReview id. It chains together publisher APIs, open-access indexes and repositories in a fallback sequence, checks that what came back is really the requested article, and can be used from Python, from the command line, or by an AI agent through an MCP server or a Claude Code skill.

Video tutorial: https://youtu.be/fTtc4QWMYzE

---

Features

• Multiple retrieval methods – Elsevier, Wiley and Springer Nature APIs, CrossRef TDM links, Unpaywall, Europe PMC, OSTI (accepted manuscripts of DOE-funded articles), Semantic Scholar's open-access index, arXiv, ChemRxiv, bioRxiv/medRxiv, Zenodo, MDPI's CDN, and direct scraping for publishers that lack easy APIs (PLOS, eLife, Cambridge, APS).
• Automatic fallback logic – The package selects the best method based on the DOI's publisher; if one fails, the next is tried automatically.
• Verification – Every PDF is checked against the article's title, and supporting-information files and abstract-only records are rejected, so a returned file is the paper you asked for.
• Honest results – Each download reports which route produced the file and carries a note when it is a preprint or accepted manuscript rather than the publisher's version of record.
• Configurable tool order – Per-publisher method sequences are configurable; defaults cover most major publishers and preprint servers.
• Batch downloads – Concurrent downloads with publisher rate limits enforced, a tqdm progress bar, and file logs that record which tool succeeded or why an identifier failed.
• Easy API-key management – Store credentials via environment variables or the interactive fulltext-config script.
• Four entry points, one engine – Python API, fulltext-download CLI, fulltext-mcp server for any MCP client, and a Claude Code plugin with a ready-made skill.

---

1. Installation

pip install fulltext-article-downloader

Optional extras enable additional routes and the MCP server:

tls and aps both work by making a request look more like your own browser than a script: tls matches Chrome's TLS fingerprint for hosts that turn away plain HTTPS clients, and aps reuses the APS session cookie you are already signed in with. Neither opens anything you are not licensed for, but both go a step beyond a plain API client, so they are worth checking against your institution's agreements before you enable them. A default install has neither; all includes them.

pip install "fulltext-article-downloader[mcp,tls]"

For development, clone the repository and run pip install -e ".[dev,all]". Make sure to configure your installation afterwards (see next section).

2. Configuration (API keys & email)

All credentials are optional; each one unlocks a route. Without keys, and off campus, expect open-access papers, preprints and DOE-funded manuscripts to work and most paywalled articles to fail.

UNPAYWALL_EMAIL is sent only to Unpaywall and Crossref, which ask API users for a contact address; Crossref serves requests that carry one from its faster "polite" pool. Keys and the email stay on your machine.

Set these environment variables or run the interactive helper:

fulltext-config

The script stores keys in ~/.fulltext_keys, which are loaded automatically on import. If a required key is missing, the corresponding tool is skipped and the downloader falls back to other methods. Publisher keys return paywalled content only when the key or the network is entitled; the package detects truncated or abstract-only responses and moves on.

---

3. Usage

Identifiers

Any of these forms is accepted everywhere an identifier is expected:

• DOI: 10.1021/jacs.3c13302, https://doi.org/10.1021/jacs.3c13302, doi:10.1021/jacs.3c13302
• arXiv: 2310.19377, arXiv:2310.19377, https://arxiv.org/abs/2310.19377, cond-mat/9712061, 10.48550/arXiv.2310.19377
• PubMed Central: PMC6561843
• OpenReview: fNyXCCZ0g6

Command-line interface (CLI)

fulltext-download <ID> [<ID> ...] [-o DIR] [--tools a,b,c] [--workers N] [--no-check] [--log-file FILE]

fulltext-download 10.1371/journal.pone.0171501 -o papers
fulltext-download 10.1021/jacs.3c13302 arXiv:1710.10324 PMC6561843 -o papers --workers 4

The command prints one JSON object per identifier and exits 0 only when every download succeeded:

[
 {
  "identifier": "10.1021/jacs.3c13302",
  "success": true,
  "path": "papers/10.1021_jacs.3c13302.pdf",
  "source": "osti",
  "note": "OSTI accepted manuscript, not the publisher's version of record",
  "error": null,
  "attempts": [],
  "supplements": []
 }
]

The original form fulltext-download <DOI> <OUTPUT_DIR> [<FILENAME>] still works.

Python API

fetch returns the same structure the CLI prints; fetch_many downloads a list concurrently:

from fulltext_article_downloader import fetch, fetch_many

r = fetch("10.1371/journal.pone.0171501", "papers")
if r["success"]:
    print(r["path"], r["source"], r["note"])
else:
    print(r["error"])           # every route tried, with its reason

rs = fetch_many(["10.1002/advs.201900808", "10.48550/arXiv.2207.03928"], "papers", workers=4)

The earlier functions are unchanged: download_article(doi, output_dir, ...) returns the path or raises, and bulk_download_articles(dois, output_dir, log_file=..., sleep=..., workers=...) returns a dict of paths or "ERROR: ..." strings with a progress bar.

Options shared by all of them: output_filename (single download), tools (a list of route names that overrides the publisher default), log_file (append a download log), and for fetch also check=False to skip verification and skip_existing=False to re-download a file that is already present.

Supplementary files

fetch(..., supplements=True), fetch_many(..., supplements=True), fulltext-download --supplements and the MCP tools' supplements=true also fetch the article's supplementary files (supporting information, data tables, videos). They are separate files from separate places, so they are saved next to the article as <name>_si1.pdf, <name>_si2.xlsx, ... and listed in the result's supplements. Sources: the ChemRxiv and Elsevier APIs, Europe PMC's supplement bundle for PMC articles, and otherwise the article's landing page (Springer Nature, Wiley, bioRxiv, ACS, RSC, PLOS and others link them there; most publishers serve supplements without a subscription). Off by default because it costs one more request per article; a missing supplement never fails the download.

MCP server (for agents)

fulltext-mcp exposes two tools, get_paper(identifier, output_dir="", tools=None, supplements=False) and get_papers(identifiers, output_dir="", max_workers=4, supplements=False), returning the structure shown above. It needs the mcp extra and reads the same keys as the CLI. FULLTEXT_OUTPUT_DIR sets where files go when a call gives no output directory (default ./papers).

Claude Code:

claude mcp add fulltext-article-downloader -- uvx --from "fulltext-article-downloader[mcp]" fulltext-mcp

Any other MCP client, in its server configuration:

{
  "mcpServers": {
    "fulltext-article-downloader": {
      "command": "uvx",
      "args": ["--from", "fulltext-article-downloader[mcp]", "fulltext-mcp"],
      "env": { "UNPAYWALL_EMAIL": "you@example.org", "FULLTEXT_OUTPUT_DIR": "/abs/path/papers" }
    }
  }
}

uvx fetches the package from PyPI into an isolated environment on first use, so nothing needs to be installed beforehand. The server is also listed in the official MCP registry as io.github.computron/fulltext-article-downloader.

Claude Code plugin and skill

This repository is also a Claude Code plugin. It installs a skill that teaches Claude when and how to use fulltext-download, plus the MCP server above:

/plugin marketplace add computron/fulltext-article-downloader
/plugin install fulltext-article-downloader@fulltext-article-downloader

The skill alone (no MCP) is enough inside Claude Code: it runs the CLI and reads the JSON. The MCP server is for clients that cannot run shell commands, and for other agent frameworks.

---

4. Failures and tools

Failure examples

Many articles are not open-access, and publishers explicitly restrict or discourage text and data mining. This example is expected to FAIL:

fulltext-download 10.1109/GROUP4.2007.4347715 -o papers

The error field lists every route tried and why it failed. A failure almost always means the article is paywalled with no open-access copy, or that a host blocked automated access.

If you see an error like Failed to load APS cookies, please make sure you are running Python using an application that has full disk access.

Verification

A file that starts with %PDF- is not proof of anything: open-access indexes occasionally map a DOI to an unrelated document, and a supporting-information file can carry the article's title. After every route the package extracts the first two pages and checks that the Crossref title is there, rejects files that open with a Supporting Information heading, and rejects Elsevier XML records that contain no body text. A rejected file is deleted and the next route is tried. --no-check / check=False turns this off.

Methods and fallback logic

The tool is composed of multiple sub-tools intended to support various publishers. The tool order depends on the publisher.

arXiv, PMC and OpenReview ids go straight to arxiv, pmc and openreview. pmc fetches Europe PMC's PDF render and, when that endpoint refuses (it throttles hosts that ask for many articles in a row), the full-text JATS XML from Europe PMC's REST service or from NCBI's efetch, which also serves author manuscripts (NCBI_API_KEY is optional and raises NCBI's rate limit).

What the less obvious tools do:

• elsevier asks for the PDF first and keeps it only when the API key is entitled to the full PDF; Elsevier otherwise answers with HTTP 200 and a PDF containing only the article's first page, so the tool discards that and takes the full-text XML instead, writing a markdown rendering next to it. Abstract-only XML is rejected. The returned path always carries the extension of the format actually written.
• crossref_tdm tries every full-text link registered with Crossref, including the unspecified content-type links that APS, ACS and RSC use; they return the PDF on an entitled network.
• europepmc downloads PMC-hosted open-access articles and the repository copies (usually author manuscripts) listed in Europe PMC's record for the DOI. When the record says the article's full text is in PMC (author manuscripts deposited under funder mandates), it takes the PMC copy: the PDF render, or the JATS XML when that endpoint refuses.
• osti queries the OSTI API for the DOI and downloads the accepted manuscript that DOE-funded articles receive on osti.gov about a year after publication. It needs no credentials; records still under embargo have no full text and the tool moves on.
• semantic downloads the open-access PDF Semantic Scholar has indexed for the DOI, mostly arXiv and institutional-repository copies of subscription articles.
• unpaywall and semantic also accept repository landing pages: when an index lists the page rather than the file (HAL, DSpace, Columbia Academic Commons), the tool reads the citation_pdf_url tag the page carries for Google Scholar and downloads that.
• chemrxiv goes through the Cambridge Open Engage API, which works from hosts that chemrxiv.org itself blocks.
• biorxiv asks the bioRxiv API which server (bioRxiv or medRxiv) holds the DOI and which version is current, then fetches that PDF. Requests are paced (the site answers 429 with a 100 s Retry-After after a burst) and the transient 503s it returns are retried; when the API itself is throttling, the unversioned URL, which redirects to the current version, is tried on both servers.
• mdpi builds the article's path on MDPI's CDN (mdpi-res.com) from the Crossref record; www.mdpi.com itself refuses requests from cloud-provider address ranges.
• zenodo downloads the PDF attached to a Zenodo record (10.5281 DOIs); a concept DOI resolves to the latest version.
• wiley is also used for society journals hosted on Wiley Online Library (AGU and others): Crossref lists their full-text links on wiley.com, and the TDM API serves them. Likewise elsevier is added for journals whose Crossref links point at Elsevier, and it closes the default list because Elsevier's API also serves society journals it distributes (ASH's Blood) under their own publisher name; a non-Elsevier DOI costs one quick 404.
• semantic also downloads the arXiv version named in a record's externalIds when the record lists no PDF, and with SEMANTIC_SCHOLAR_API_KEY searches by title for a second record of the same paper (the search endpoint answers 429 without a key). Requests are spaced one second apart, Semantic Scholar's limit.

Downloads send a browser User-Agent, retry with a plain one for repositories that serve PDFs only to non-browser clients, and, with the tls extra, once more with a Chrome TLS fingerprint. Wiley requests are paced to 30 per 10 minutes (the published limit is 60, but the API answers HTTP 500 from about 35 on), Crossref requests to its polite-pool limit, and transient 429/5xx responses are retried once.

Routes that return preprints or accepted manuscripts (arxiv, chemrxiv, biorxiv, osti, semantic, repository copies from unpaywall and europepmc) set note in the result. See PUBLISHER_TOOL_MAP in downloader.py.

Tip – Scraping-based methods (springerpdf, elife, cambridge, etc.) can break if sites change layout or due to access limits; favour official APIs and Unpaywall for large-scale downloads.

Customising the tool sequence

from fulltext_article_downloader import fetch
fetch("10.1017/S0885715624000484", "papers", tools=["unpaywall", "cambridge"])  # override default order

Override at runtime:

from fulltext_article_downloader import PUBLISHER_TOOL_MAP
PUBLISHER_TOOL_MAP["Elsevier BV"] = ["elsevier", "unpaywall"]

---

5. License and Disclaimer.

BSD 3-Clause. See the LICENSE file.

Use this tool only for content you are legally entitled to access. Respect publisher terms and copyright laws. It does not use Sci-Hub or any similar site, and retrieves only what your own subscriptions, API keys and open-access licenses already permit; the optional tls and aps routes present your requests as a browser would, as described under Installation. The authors are not responsible for misuse.

Speed-coded by computron on vibes (ChatGPT 4.0) and caffeine; extended by Xu Huang with Claude.