Installation

Installation

Baca dalam bahasa Indonesia

Claude Desktop — the easy path

Not sure which Claude you have? If you use Claude in a browser or a desktop application, this is your section. If you type claude into a terminal, skip to Claude Code.

1. Download the server you want. Each link saves the file straight to your computer:

   Server	What it does	Download
   scholar-nulis	search literature, verify citations	scholar-nulis-0.8.0.mcpb
   zotero-nulis	read your own Zotero library	zotero-nulis-0.7.0.mcpb
   scr-toolkit-nulis	deterministic scoping-review checks	scr-toolkit-nulis-2.0.0.mcpb

2. Double-click the file. Claude Desktop opens an install window. (Alternative: Settings → Extensions, then drag the file in.)

3. Fill in the configuration fields if you need them — all optional for scholar-nulis, and scr-toolkit-nulis has none at all.

4. Click Install, then quit Claude Desktop completely and reopen it.

Closing the window is not enough; servers are only re-read when the application actually restarts. On macOS, make sure the icon is gone from the Dock (⌘Q).

Nothing needs to be installed first. Claude Desktop ships its own Node.js.

Upgrading from the old names

Every tool name now begins with nulis_ — nulis_search_arxiv, nulis_zotero_search_items, nulis_pdf_integrity, and so on. This is a breaking change, and it exists because Claude Desktop presents tool names flat: there is no per-server namespace like Claude Code's mcp__scholar__. Two extensions that both register search_arxiv therefore collide, and which one wins is not predictable. Measured on one real machine, this repo's servers shared nine tool names with unrelated MCP servers already installed; after the prefix, zero.

You do not need to change how you talk to Claude — it reads the tool list itself. Only scripts or written instructions that name a tool literally need updating.

These servers were previously called scholar-paper-search, zotero-mcp, and scr-toolkit. Claude Desktop identifies an extension by the name in its manifest, so a renamed server is a different extension as far as it is concerned — installing the new one does not replace the old one, and you will end up with both, exposing duplicate tools.

Remove the old entries first in Settings → Extensions, then install these. The suffix exists precisely to stop this kind of ambiguity: scholar is a common enough name that a second, unrelated MCP server may well already be using it.

Claude Code — build from source

git clone https://github.com/nulis-not-just-writing/mcp-writing-toolkit.git
cd mcp-writing-toolkit

# scholar
cd scholar-nulis && npm install && npm run build && cd ..
claude mcp add scholar -- node "$PWD/scholar-nulis/dist/index.js"

# zotero
cd zotero-nulis && npm install && npm run build && cd ..
claude mcp add zotero -- node "$PWD/zotero-nulis/dist/index.js"

# scr-toolkit-nulis — no dependencies, so no npm install and no build
claude mcp add scr-toolkit -- node "$PWD/scr-toolkit-nulis/server/index.js"

To pass configuration, use -e before --:

claude mcp add scholar \
  -e CONTACT_EMAIL=you@university.edu \
  -e SCOPUS_API_KEY=xxxxx \
  -- node "$PWD/scholar-nulis/dist/index.js"

Check the result with /mcp inside a Claude Code session.

Configuration

scr-toolkit-nulis has no configuration fields at all — install and use. The other two read theirs only from environment variables; in Claude Desktop, manifest.json populates them from the extension form, so you never touch them.

There is no .env file. No server here reads one. If you find older instructions telling you to copy .env.example, they belong to the previous Python-based pack and no longer apply.

scholar

Field / variable	Required	Effect when set
CONTACT_EMAIL	no	Joins the Crossref & OpenAlex polite pool (higher quota) and adds the Unpaywall route for open-access lookup, which additionally reports each PDF's licence. Any working address; no registration.
S2_API_KEY	no	Raises the Semantic Scholar rate limit. Free at semanticscholar.org/product/api.
DOWNLOAD_DIR	no	Where PDFs are stored. If empty: ~/Downloads, then the system temp folder if ~/Downloads is not writable.
SCOPUS_API_KEY	no	Switches on nulis_search_scopus, nulis_scopus_abstract, nulis_scopus_export_csv, nulis_elsevier_status.
SCIENCEDIRECT_API_KEY	no	Switches on nulis_sciencedirect_fulltext. Leave empty if your Scopus key already covers it.
ELSEVIER_INSTTOKEN	no	Institutional token from your librarian or licence admin. Needed only when off-campus access is refused with 401/403.

Elsevier keys are free to register with an institutional account at dev.elsevier.com. The quota is tied to your campus subscription.

zotero

Field / variable	Required	Effect
ZOTERO_LOCAL	no	true is the default — talks to the Zotero app on this computer.
ZOTERO_API_KEY	Web API mode	Key from zotero.org/settings/keys.
ZOTERO_LIBRARY_ID	Web API mode	Your user ID, on the same page.
ZOTERO_LIBRARY_TYPE	Web API mode	user or group.

Local mode needs one step inside Zotero: open Zotero → Settings → Advanced → tick "Allow other applications on this computer to communicate with Zotero". Without it the server will report that Zotero is unreachable.

The Zotero application must be running when a tool is called.

Keeping keys safe

Fields marked sensitive in the extension window are stored by Claude Desktop in your operating system keychain, not as plain text.

If you build the variant bundle with an embedded key for a team (scholar-nulis/build-team-bundle.sh), remember that the resulting file carries that key inside its manifest. That bundle must never be uploaded anywhere public; this repo's .gitignore already refuses it through the *-api.mcpb pattern, and build-mcpb.sh refuses to publish any bundle whose manifest carries a literal key.

Rebuilding bundles

./build-mcpb.sh                 # all three
./build-mcpb.sh zotero-nulis    # just one

Output goes to dist/<name>-<version>.mcpb. The script stops and deletes any bundle that fails a gate — a broken bundle left lying in dist/ will eventually be distributed by accident.

The gates:

• The manifest.json version must match package.json (for servers that have one).
• The version the server announces over an MCP handshake must match its manifest. This gate runs the packed server and asks it directly — a version hardcoded in source is exactly the kind of drift a file-level check misses, and it really did happen to all three servers in this repo.
• NOTICE.md must be inside the bundle — attribution for bundled libraries has to travel with every copy, and a .mcpb downloaded on its own does not carry the repo.
• No node_modules/, src/, .env, or junk files.
• No literal credential values in mcp_config.env.

Servers without a package.json (like scr-toolkit-nulis) are packed as-is with no build step; the gate then confirms the entry_point named in the manifest actually exists.

---

← Back · scholar · zotero · scr-toolkit · FAQ