Academic Search Skill

Search and download research papers from the terminal. Wraps paper-search-mcp (arXiv, PubMed, bioRxiv, medRxiv, Google Scholar) + arxiv-dl (fast arXiv PDF downloads).

SKILL_DIR below = this skill's own directory — skills/tool-academic-search/ in a clone, or ~/.claude/skills/tool-academic-search/ in the per-machine store. Reference the script with that absolute path.

Isolation (important)

The underlying tools live in a dedicated venv at %LOCALAPPDATA%/eos/envs/academic/ — NOT the EmptyOS daemon env. They drag in rich>=14 / starlette>=1.0 which conflict with the daemon's pins, so they were deliberately isolated (per .claude/rules/environment.md § user-home Python envs). academic.py re-execs itself under the venv python automatically, so you can call it with a plain python — no activation needed.

Setup (one-time, already done)

uv venv --python 3.13 "%LOCALAPPDATA%/eos/envs/academic"
uv pip install --python "%LOCALAPPDATA%/eos/envs/academic/Scripts/python.exe" arxiv-dl paper-search-mcp

Usage

1. Search

python "<SKILL_DIR>/academic.py" search "diffusion models" --source arxiv -n 5
python "<SKILL_DIR>/academic.py" search "crispr gene editing" --source pubmed
python "<SKILL_DIR>/academic.py" search "battery degradation" --source all -n 3
python "<SKILL_DIR>/academic.py" search "graph neural network" --json     # machine-readable

• --source: arxiv (default) | pubmed | biorxiv | medrxiv | scholar | all
• -n: max results per source (default 8)
• --json: emit a JSON array (id, title, authors, date, source, url, doi) — use this when you need to parse results to pick one to download.

Per-source failures are soft (printed to stderr) so --source all never dies on one bad backend.

2. Download a PDF

python "<SKILL_DIR>/academic.py" download 1706.03762 --dir ./papers --pdf-only
python "<SKILL_DIR>/academic.py" download https://arxiv.org/abs/1512.03385 --dir ./papers
python "<SKILL_DIR>/academic.py" download 42318429 --source pubmed --dir ./papers

• arXiv targets route through arxiv-dl (fast, parallel, clean filenames).
• --source pubmed|biorxiv|medrxiv use the respective backend; pass a PMID or DOI.
• --pdf-only (arXiv) skips the accompanying notes file.

3. Read a paper's full text (stdout)

python "<SKILL_DIR>/academic.py" read 1706.03762 --source arxiv

Downloads then extracts the paper text. Costs context — prefer this only for short papers or when you need the content in-conversation. For a long paper, download it then use tool-pdf-reader's dump mode (token-cheaper).

Workflow: find → download → file into KB

1. Search and show the user candidates:

   python "<SKILL_DIR>/academic.py" search "<topic>" --source all -n 5
   

2. On their pick, download the PDF to a scratch or vault attachments dir:

   python "<SKILL_DIR>/academic.py" download <id> --dir "{vault}/99_Attachments/papers" --pdf-only
   

3. To file it into the EmptyOS KB, hand the downloaded PDF to vault-source-digest (creates the reference landing note + clause/case notes). To just read it locally, use tool-pdf-reader.

Sources covered

Source	search	download	read	id format
arXiv	✓	✓ (arxiv-dl)	✓	2606.20554 / abs URL
PubMed	✓	✓	✓	PMID
bioRxiv	✓	✓	✓	DOI
medRxiv	✓	✓	✓	DOI
Google Scholar	✓	—	—	(search only)

Google Scholar is search-only (no stable PDF endpoint); use it to discover a title, then download from arXiv/PubMed if available.

When NOT to use this

• The PDF is already on disk → tool-pdf-reader.
• Goal is digesting a standard/textbook into the KB → vault-source-digest (hand it the downloaded PDF).
• You need a Google Scholar PDF directly → not supported; find the paper on arXiv/PubMed instead.