snu-library
by SleepingDawn
README.md
# SNU Research Agent — MCP + Skill
Local scholarly full-text evidence for Codex and Claude. Native search discovers papers; this package accesses permitted PDFs through SNU and returns page-grounded text. It is not a global full-text search index.
## Install
```sh
git clone https://github.com/SleepingDawn/snu-research-agent.git
cd snu-research-agent
python3 -m venv .venv
.venv/bin/python -m pip install -e .
.venv/bin/python -m playwright install chromium
.venv/bin/scholarly-access-agent setup
.venv/bin/scholarly-access-agent doctor
```
Connect MCP using `examples/claude-mcp.json` or `examples/codex-mcp.toml`, replacing the executable path. Claude Code: `claude mcp add --scope user snu-library -- /absolute/repo/.venv/bin/scholarly-access-agent mcp-server`. For Claude account Skills upload `dist/snu-library-scholarly-access.zip`. Restart clients after changing tools. A cloud-only session needs a local MCP connection to use local access.
## Authentication and research
Run `scripts/서울대 로그인.command` for user-assisted SNU sign-in. Passwords are never loaded or auto-entered. Routine research uses native search and CLI/MCP only, with headless downloads. Session renewal alone opens a window. Session files stay outside the repo, owner-only and unencrypted.
The canonical skill is `skills/snu-library-scholarly-access/SKILL.md`; references separate workflow, DFT/MLIP criteria, evidence format and authentication. Installed copies add a local executable hint, excluded from the upload ZIP.
Tools: `doctor`, `check_url`, `proxy_url`, `discover_pdfs`, `download_pdf`, `read_pdf`, `add_domain`. MCP download operations always run headless. CLI and MCP share the same reading/downloading implementation. Doctor checks session presence, not validity. Subscription access does not establish external AI permission; see https://lib.snu.ac.kr/find/databases_openaccess/tdm/ . Scanned PDFs require OCR or visual review. For `read_pdf`, choose `terms` OR `pages`; never both. `pages_scanned` reports search coverage, not pages fully reviewed by the agent.
## Verify
```sh
.venv/bin/python -m unittest discover -s tests -v
.venv/bin/python -m pip check
```
CI uses synthetic fixtures and needs no SNU credentials. Before private GitHub upload, inspect staged files and history; exclude sessions, credentials, PDFs, reports and generated packages. Runtime data remains local even when this source repository is private.
## Provenance
Based on https://github.com/onetwothr1/SNU-library-scholarly-access-skill at `cad5e9a5e9ab259642b5dee13d6022427e56bd03`. Upstream README declares MIT. The original MIT notice, including its copyright attribution, is retained in `LICENSE`, recovered verbatim from upstream commit `b5d7927` (it was absent at the review revision). No ownership claim is made over upstream code.
Design references (consulted 2026-11-17):
- https://github.com/K-Dense-AI/scientific-agent-skills/tree/main/skills/literature-review
- https://github.com/K-Dense-AI/scientific-agent-skills/tree/main/skills/research-lookup
- https://liyux3.github.io/scholar-mcp/research-workflows/
- https://github.com/Future-House/paper-qa
Their workflow patterns informed original instructions here; their code was not vendored.
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessNo recent community issues