Paper library

When a conversation ends, the research stays in the library. It is a local SQLite database with full-text search. Each record keeps a stable identifier, the full author list and abstract, your notes and tags, the PDF path, and the IDs of the matching Zotero or Notion entries. Agents read and write it through paper_library in the research profile, and the command line shows the same records.

A citation graph drawn from a real local collection.

Collections, notes and tags

Papers are kept in collections, default unless you name another. paper_library supports these actions.

Action	What it does
save or add	Resolve one or more identifiers and store them with optional notes and tags
get	Return one saved paper with its notes
list	List a collection, with notes, tags and PDF paths
search	Full-text search over saved titles, abstracts, authors and notes
update	Replace the notes or tags of one paper
remove	Permanently delete one paper from a collection
collections	List collections and their sizes
export	Write the collection as an Obsidian-compatible Markdown vault

Saving the same paper again adds new notes without duplicating them. Only update replaces a note. download_paper attaches the PDF to the matching record, so a paper and its file stay together.

A request such as save these three to the retrieval collection, tag them baseline and note why each matters is enough for the agent to call save with the right identifiers.

Where files are kept

~/.scholar-mcp/
├── papers/                persistent PDFs
├── kb/
│   ├── library.sqlite3    authority + FTS5 + sync state
│   └── *.jsonl            compatibility snapshots
└── vault/                 Markdown projections and wikilinks

SCHOLAR_DATA_DIR moves the whole tree. SCHOLAR_KB_DIR and SCHOLAR_DOWNLOAD_DIR move the library and the PDFs separately. Older JSONL collections move into SQLite automatically and stay as snapshots.

Command line

The same records are available in a terminal. Command line lists every subcommand.

scholar-mcp library doctor
scholar-mcp library list --collection retrieval
scholar-mcp library search "contrastive pretraining" --collection retrieval

Connectors

Each connector records external IDs, versions and content fingerprints, so unchanged papers are not published twice.

Obsidian

Obsidian needs no login. Each paper becomes one Markdown note, plus an index for the collection, linked with wikilinks and written straight into the vault.

scholar-mcp library export obsidian --collection retrieval --path /path/to/vault

--link-citations fetches reference lists and links notes along real citations. With SCHOLAR_OBSIDIAN_VAULT set, the vault stays in sync, and every save, update and download refreshes it.

Zotero

Zotero receives items, collections and tags, with a separate Scholar research note under each paper. Set ZOTERO_API_KEY and ZOTERO_LIBRARY_ID, and for a group library or the local API also ZOTERO_LIBRARY_TYPE or ZOTERO_API_BASE.

# Dry run: prints the plan
scholar-mcp library sync zotero --collection retrieval

# Performs the writes
scholar-mcp library sync zotero --collection retrieval --apply

--zotero-collection-key targets an existing Zotero collection.

Notion

Notion receives a reading list with a readable summary and editable research notes, even when the database has only a title column. Set NOTION_API_KEY and NOTION_DATA_SOURCE_ID.

# Dry run: prints the plan
scholar-mcp library publish notion --collection retrieval

# Performs the writes
scholar-mcp library publish notion --collection retrieval --apply

Your own Notes property is never overwritten.

Dry run by default

Zotero sync and Notion publish only print a plan until you add --apply. Obsidian export writes local files directly.

Your edits are read back

Edit notes where you read. After an export, write in the Research notes section or under Notes in the Markdown file, or edit the research note Scholar created in Notion. The next library read, whether get, list, search or update, picks up your edits first, and the agent continues from them.

When both sides changed, your version wins, even if you cleared the note. An untouched Markdown file never rolls back a newer draft in the library. Conflicts and read errors are listed in the response as note_conflicts and note_errors.

Connector credentials can be kept in the system’s encrypted credential store. Command arguments and environment variables take precedence, and credentials are never written to the library.