Endnote MCP

EndNote MCP

Connect your EndNote reference library to Claude AI. Search references, read PDFs, format citations, find related papers, and generate bibliographies — all directly in Claude Desktop conversations.

What It Does

Once set up, you can ask Claude things like:

• *"Search my library for social capital and Bourdieu"*
• *"Find papers about how organisations deal with uncertainty"* (semantic search)
• *"Find references related to paper #3844"*
• *"Give me the APA citation for reference #1234"*
• *"Generate a bibliography for references 12, 45, 78, 102"*
• *"Export references 12, 45, 78 as BibTeX"*
• *"Read pages 5-7 from that Smith et al. paper"*

Claude searches your local library — nothing is uploaded to the cloud beyond the normal conversation.

How It Works

EndNote Library → XML Export → endnote-mcp index → SQLite Database (FTS5 + Embeddings)

↕ Claude Desktop ← MCP Server (12 tools)

Your references and PDF text are indexed into a local SQLite database with full-text search. Optionally, AI embeddings enable semantic search — finding papers by meaning, not just keywords. When semantic dependencies are installed, new references are automatically embedded during indexing. Claude connects through the Model Context Protocol (MCP).

Requirements

• EndNote 20+ (any edition)
• Claude Desktop app

The bundle installer below has no other prerequisites — Claude Desktop manages Python and dependencies. The CLI install path additionally needs Python 3.10+ and uv (or pip).

Install (Desktop Bundle — Recommended)

The fastest install for non-engineers. No Python or terminal required.

1. Export your library from EndNote. File → Export → choose XML format → save somewhere like Desktop.
2. Download endnote-mcp.mcpb from the latest release.
3. Double-click the `.mcpb` file (or in Claude Desktop: Settings → Extensions → Install Extension…).
4. In the install dialog, point Claude Desktop at:

• Your EndNote XML file (from step 1)
• Your PDF Attachments Folder — usually .Data/PDF/ next to your .enl file, or PDF/ inside an .enlp package.

1. Click Install. Claude Desktop will start indexing your library automatically — metadata is searchable within seconds, PDF text and semantic embeddings finish in the background.

That's it. Try asking Claude *"Search my library for grounded theory"*.

The bundle keeps its own configuration (config-mcpb.yaml), so installing it will not disturb an existing CLI setup. Both default to the same database, so if you already indexed with the CLI the bundle reuses that index rather than building a second copy.

Building the bundle yourself

Maintainers (or anyone working from source) can build it with:

bash scripts/build_mcpb.sh

→ dist/endnote-mcp.mcpb

The build is a single cross-platform .mcpb file (~36 KB). Dependencies resolve at install time via the uv runtime bundled with Claude Desktop.

The format was called DXT and used the .dxt extension until it was renamed upstream to MCPB (MCP Bundles). See modelcontextprotocol/mcpb.

Install (CLI / from PyPI)

Prefer the terminal? This path gives you the endnote-mcp CLI for indexing and embedding.

1. Install

# With uv (recommended)

uv tool install endnote-mcp

Or with pip

pip install endnote-mcp

2. Export your library from EndNote

In EndNote: File → Export → choose XML format → save to a convenient location (e.g., Desktop).

3. Run the setup wizard

endnote-mcp setup

The wizard will:

• Auto-detect your XML export and PDF directory
• Create the configuration
• Index your library
• Configure Claude Desktop automatically

4. Restart Claude Desktop

Quit and reopen Claude Desktop. You'll see "EndNote Library" in your MCP connectors.

Claude Code Plugin and Agent Skill

The repo includes a skill, endnote-library, that teaches an agent how to work with your library: cite only records the tools return, take page numbers from the printed page rather than the PDF position, check quotations word for word, and verify a draft's references one by one.

Claude Code. Install the CLI and index your library first (steps 1–3 above), then run:

/plugin install endnote --marketplace gokmengokhan/endnote-mcp

On Claude Code before 2.1.275, run /plugin marketplace add gokmengokhan/endnote-mcp, then /plugin install endnote@endnote-mcp. The plugin adds the skill and starts the server with endnote-mcp serve, so that command must be on your PATH. If you already added the server with claude mcp add and the same command, Claude Code connects it only once; if yours uses another command or path, remove it with claude mcp remove endnote to avoid duplicate tools.

Claude Desktop. Zip the plugin/skills/endnote-library/ folder and upload it under Customize > Skills.

Other agents. Agents that read Agent Skills (SKILL.md folders) can use the same skill: copy plugin/skills/endnote-library/ into the agent's skills directory.

Semantic Search (Optional)

Enable meaning-based search that finds references even when they use different terminology than your query. For example, searching *"how companies prepare for uncertain futures"* finds papers on scenario planning and strategic foresight.

# Install semantic search dependencies

pip install endnote-mcp[semantic]

Generate embeddings (~3 min for 4,000 references)

endnote-mcp embed

This uses the lightweight all-MiniLM-L6-v2 model locally — no API keys needed, everything stays on your machine.

Commands

Tools Available to Claude

Adding New References

When you add new references to your EndNote library:

1. Re-export XML from EndNote (overwrite the same file)
2. Either:

• Run endnote-mcp index from a terminal, or
• Ask Claude: *"Rebuild my library index"*

Indexing is incremental — it only processes new references and PDFs, not the entire library again. If semantic search is installed, new references are automatically embedded.

Performance

Configuration

Config is stored at:

• macOS: ~/Library/Application Support/endnote-mcp/config.yaml
• Windows: %APPDATA%/endnote-mcp/config.yaml
• Linux: ~/.config/endnote-mcp/config.yaml

endnote_xml: /path/to/your/library.xml

pdf_dir: /path/to/your/Library.Data/PDF db_path: /path/to/library.db # auto-set by setup max_pdf_pages: 30 # max pages per read request search_notes: false # include the EndNote Notes field in searches

Notes fields

EndNote's Research Notes — where you keep your own summaries and judgments — is indexed and searched by default, ranked just below keywords and above the abstract.

The plain Notes field is indexed but excluded from matching, because reference managers use it as a dumping ground on import: author affiliations, email addresses, thesaurus terms. Set search_notes: true to include it. The setting takes effect immediately — no re-index needed.

Citation Styles

Five built-in styles:

• APA 7th — get_citation(rec_number=42, style="apa7")
• Harvard — style="harvard"
• Vancouver — style="vancouver"
• Chicago (Author-Date, 17th ed.) — style="chicago"
• IEEE — style="ieee"

Use get_bibliography to format multiple references at once, sorted by author or year.

BibTeX Export

Export references as BibTeX entries for use in LaTeX:

• Ask Claude: *"Export references 42 and 100 as BibTeX"*
• Or use the get_bibtex tool directly with comma-separated record numbers

Generates complete entries with proper entry types (@article, @book, @inproceedings, etc.), cite keys, and all available fields.

Troubleshooting

"No configuration found" — Run endnote-mcp setup

"XML file not found" — Re-export from EndNote: File → Export → XML format

"PDF not found" — Check that pdf_dir in your config points to the correct .Data/PDF directory

Search returns no results — Run endnote-mcp index to rebuild the database

Claude Desktop doesn't show the tool — Run endnote-mcp install, then restart Claude Desktop

"Semantic search is not available" — Run pip install endnote-mcp[semantic] then endnote-mcp embed

Client reports the server "failed to connect" / `CONNECTION_CLOSED` — The client hides the underlying error. Run echo '' | endnote-mcp serve to see the real traceback (the JSON decode error it prints afterwards is just the empty input, not a fault).

Releasing

Publishing runs on a tag push via .github/workflows/release.yml, which uploads to PyPI and the MCP Registry using OIDC — no API tokens are stored in the repository.

The version lives in seven places, and the registry rejects a server.json that does not exactly match the published PyPI release. Bump all of them, then verify before tagging:

# pyproject.toml, server.json (x2), CITATION.cff, the README citation

block, mcpb/manifest.json and mcpb/pyproject.toml

python scripts/check_version.py v