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