paper-fetch: A Zero-Dependency, Agent-Native Tool for Legal Open-Access PDF Retrieval

Overview

paper-fetch is an open-source Agent Skill maintained by Agents365-ai that solves a foundational problem in AI-driven research: reliably obtaining legal open-access (OA) PDFs for AI agents. Given a DOI, it returns a PDF envelope to the caller. The project is MIT-licensed, currently holds 83 stars and 6 forks, and implements the entire toolchain using only the Python standard library (Python 3.8+).

Key Design Points

• Zero-dependency philosophy: Implemented entirely with Python's standard library. No pip install required, making it portable to containers, sandboxes, and offline agent environments where dependency installation may fail.
• Six-source fallback chain queried in priority order:
1. Unpaywall — highest cross-disciplinary coverage via Crossref. 2. Semantic Scholar — uses openAccessPdf and externalIds cross-references. 3. arXiv — physics, math, CS, statistics, economics, EE preprints. 4. PubMed Central (PMC) — biomedical full-text repository. 5. bioRxiv / medRxiv — DOI prefix 10.1101/ handling. 6. Sci-Hub mirrors — last-resort fallback, disabled via PAPER_FETCH_NO_SCIHUB=1. When no legal OA source is available, the tool returns metadata rather than bypassing paywalls.
• Agent-native CLI:
• JSON envelope to stdout.
• NDJSON progress streams to stderr for batch monitoring.
• Typed exit codes: 0 success, 1 general error, 3 partial success, 4 no result.
• Idempotency-key caching (--idempotency-key) for replayable batch results.
• Schema introspection: scripts/fetch.py schema --pretty.
• TTY-aware output (human-readable vs. JSON, auto-detected).
• Self-update mechanism: Auto-runs git pull --ff-only at most once per 24 hours, throttled via a <skill_dir>/.last_update file. Fast-forward only, safe by default, can be disabled.

Security and Legal Posture

• Strictly legal OA retrieval only; refuses to bypass paywalls.
• Domain whitelist for all download URLs.
• 50 MB per-file size cap.
• SSRF defenses: rejects private IPs, non-HTTP(S) schemes, ports outside 80/443, and cloud-metadata hostnames.
• Validates %PDF magic bytes post-download to fail cleanly on publisher HTML redirects.
• Optional institutional proxy support via PAPER_FETCH_INSTITUTIONAL=1.

Competitive Comparison

| Tool | Dependencies | Agent-Native | Multi-Source | Legal OA | Batch | Self-Update | |---|---|---|---|---|---|---| | paper-fetch | None | Yes | 6 sources | Whitelisted | Yes | Yes | | unpywall | pandas | No | 1 source | Yes | No | No | | PyPaperBot | Light | No | Sci-Hub-centric | Grey area | No | No | | PyPaperRetriever | Medium | No | 3 sources | Yes | Yes | No | | hcss-utils/unpaywall | Python + Java + Bash | No | 1 source | Yes | No | No |

Distribution and Ecosystem

The skill ships as a portable SKILL.md plugin supporting multiple platforms:

• Claude Code: ~/.claude/skills/paper-fetch/
• OpenClaw / ClawHub: clawhub install paper-fetch
• Hermes Agent: ~/.hermes/skills/research/
• pi-mono: ~/.pimo/skills/paper-fetch/
• OpenAI Codex: ~/.agents/skills/paper-fetch/ (with openai.yaml sidecar)
• SkillsMP indexed.

Agents365-ai maintains a broader agent-tooling ecosystem using the same SKILL.md pattern, including podcast-skill, drawio-skill, zotero-skill, asta-skill, and excalidraw-skill.

Honest Limitations

• No institutional authentication beyond optional proxy toggle.
• DOI-centric only; no title/keyword search.
• Coverage bounded by available legal OA; honest failure is by design.
• PDF download only; no parsing, summarization, or vectorization — must be paired with downstream tools such as GROBID, LLM summarizers, or Zotero.

References

1. Agents365-ai. paper-fetch. GitHub. https://github.com/Agents365-ai/paper-fetch 2. Agents365-ai. paper-fetch project page. https://agents365-ai.github.io/paper-fetch/ 3. Awesome Skills. paper-fetch skill index. https://www.awesomeskills.dev/en/skill/agents365-ai-paper-fetch 4. unpywall. unpywall GitHub repository. https://github.com/unpywall/unpywall 5. PyPaperRetriever. Journal of Open Source Software 08135.