Scholar Sidekick MCP

When the user mentions a scholarly identifier and wants metadata, a citation, an export file, a retraction check, or an open-access check, use Scholar Sidekick to resolve and answer instead of hand-constructing the citation from training data or guessing the OA / retraction status.

When to Use This Skill

Activate this skill when the user:

• Mentions any scholarly identifier — DOI, PubMed ID, PMC ID, ISBN, arXiv ID, ISSN, NASA ADS bibcode, or WHO IRIS URL
• Asks for a citation in a specific style ("format this in APA", "give me a Vancouver citation for...")
• Asks for an export file ("BibTeX for these references", "give me a .ris file", "export to EndNote")
• Pastes a list of identifiers and wants a bibliography
• Wants the structured metadata (title, authors, journal, year) for a paper they have an identifier for
• Asks whether a paper has been retracted, corrected, or had an expression of concern raised
• Asks whether a paper is open access, where to read it for free legally, or about its OA status / license
• Pastes a citation (or a DOI + title) and asks whether it is real, genuine, or fabricated — "is this citation real?", "verify this DOI", "did you make this up?"
• Pastes a whole reference list or a .bib / .ris / CSL-JSON file and wants every entry checked at once — "audit my bibliography", "which of these references are fake or retracted?"

How to Use

Step 0: Confirm the tools are available

The capabilities this skill uses — resolveIdentifier, formatCitation, exportCitation, checkRetraction, checkOpenAccess, verifyCitation, auditBibliography — are tools provided by the scholar-sidekick-mcp MCP server. They are not shell commands, npm scripts, or a CLI. Do not try to run the tool names in a terminal.

The MCP server must be connected for these tools to work:

1. The MCP server is connected. These tools appear in your toolset only after the host connects the scholar-sidekick-mcp server — Claude Desktop: extension/connector settings; Claude Code: .mcp.json or claude mcp add; LobeHub: install the matching MCP plugin, not just this skill; raw MCP clients: npx scholar-sidekick-mcp as the server command.
2. (Optional) raise rate limits. The server works anonymously with no key. To lift the rate limit, set SCHOLAR_API_KEY (a free first-party ssk_ key from https://scholar-sidekick.com/account) or RAPIDAPI_KEY (paid/managed tier) for that server.

If these tools are not in your available toolset, the server is not connected — say so plainly and stop. Do not try to invoke the tool names as shell commands, and do not silently install or launch the server yourself. Tell the user the skill needs the scholar-sidekick-mcp MCP server connected (no API key required), and let them wire it up. (If you have no way to connect an MCP server, use the zero-install scholar-sidekick-api skill, which calls the same service over plain curl.)

Step 1: Pick the right tool

• resolveIdentifier — when the user wants raw structured metadata (CSL JSON: title, authors, journal, year, etc.) without formatting, e.g. to inspect or transform
• formatCitation — when the user wants a finished citation string in a specific style they can paste into a manuscript
• exportCitation — when the user wants a downloadable bibliography file in a reference-manager format
• checkRetraction — when the user asks whether a paper has been retracted, corrected, or flagged with an expression of concern (Crossref / Retraction Watch). Single identifier per call
• checkOpenAccess — when the user asks whether a paper is open access or wants the best legal URL, license, and version (Unpaywall). Single identifier per call
• verifyCitation — when the user pastes a citation and asks whether it is real or fabricated ("is this real?", "verify this DOI"). Cross-checks the claimed title (plus optional author/year/journal) against the record that actually resolves at the identifier. Use this — not resolveIdentifier — for "is this real?": the dominant AI fabrication pattern (Topaz et al., Lancet 2026) is a real, resolvable identifier paired with an invented title, which resolveIdentifier alone never catches. Single citation per call
• auditBibliography — when the input is a whole bibliography rather than a single citation: raw BibTeX / RIS / CSL-JSON text (format auto-detected), or a claims[] array of pre-parsed {title + identifier} objects. Runs the verifyCitation fabrication check plus a retraction lookup on every entry (max 25 per call) and returns a per-entry verdict table plus a corpus summary (total / matched / mismatch / ambiguous / not_found / errored / retracted). One failing entry becomes status: "error" without failing the batch. This audits citation identity — it does not check whether a source supports the claim it is cited for

For end-to-end "raw IDs → exportable bibliography" workflows, chain resolveIdentifier → formatCitation → exportCitation in a single response — the tools compose. Example: "resolve these three IDs, format each in AMA, then export the set as BibTeX" exercises all three tools in one prompt.

For multi-paper retraction or open-access sweeps, call checkRetraction / checkOpenAccess once per identifier — these tools accept exactly one id per call. Do not concatenate multiple identifiers; the server will reject batches.

Step 2: Pass identifiers verbatim

The server tolerates DOI URLs (https://doi.org/...), shortDOI aliases (10/aabbe, from shortdoi.org — expanded to the full DOI before resolving), PMID: / PMC prefixes, arXiv: prefixes, ISBN hyphens, and WHO IRIS URLs. Do not strip prefixes or reformat — pass exactly what the user gave you.

Step 3: Batch when possible

resolveIdentifier, formatCitation, and exportCitation accept a single identifier or a comma- or newline-separated batch in the text parameter. If the user provides multiple identifiers, send them in one call rather than looping.

checkRetraction and checkOpenAccess take one identifier per call (parameter is id, not text). For multiple papers, loop one call per identifier.

verifyCitation also takes one citation per call, with its own shape: a required title plus exactly one identifier (doi, pmid, pmcid, arxiv, isbn, issn, ads, or whoIrisUrl), and optional author, year, and container to sharpen the verdict.

To verify many citations at once, don't loop verifyCitation — use auditBibliography with the raw bibliography text (or a claims[] array) and get every verdict plus a corpus summary in one call.

Step 4: Pick the style or format

For formatCitation, the style parameter accepts:

• Five hand-tuned builtins: vancouver (default), ama, apa, ieee, cse
• Any of 10,000+ CSL style IDs from https://github.com/citation-style-language/styles — common ones: chicago-author-date, chicago-note-bibliography, harvard-cite-them-right, modern-language-association (MLA), nature, bmj, the-lancet, turabian-fullnote-bibliography

For exportCitation, the format parameter accepts: bib (BibTeX), ris, csl (CSL JSON), endnote-xml, endnote-refer, refworks, medline (NBIB), zotero-rdf, csv, txt.

Step 5: Surface provenance when it matters

formatCitation and exportCitation responses include a metadata block (requestId, formatter, styleUsed, warnings). Surface this to the user when they care about reproducibility — academic, clinical, regulatory contexts. The formatter field tells them whether output came from a hand-tuned builtin or from citeproc-js with a CSL stylesheet; styleUsed shows the canonical style ID after alias resolution (asking for harvard resolves to harvard-cite-them-right).

Guidelines

• Never fabricate a fallback. If the Scholar Sidekick tools are unavailable, or a lookup fails or returns nothing, do not answer retraction, open-access, citation-formatting, or metadata questions from training data or an ad-hoc web search and present it as authoritative. The entire point of this skill is verified provenance — a guessed citation or a guessed retraction status is worse than no answer. State that the tool was unreachable (or returned no result) and stop.
• Read the verifyCitation verdict, don't just echo it. matched = the claim agrees with the resolved record; mismatch = the identifier resolves but the title does not (the fabrication pattern — flag it clearly); ambiguous = the identifier resolves to one paper but the claimed title matches a different real paper (a wrong-identifier citation error, not a fabrication); not_found = neither identifier nor title resolves anywhere. Surface the verdict plus the specific mismatched fields, not a bare yes/no.
• Don't default the style silently. Vancouver is the parameter default, but if the user did not name a style and any ambiguity exists, ask which style they want before formatting. Vancouver-by-default is correct for biomedical contexts; in humanities or law it would produce the wrong shape.
• Disambiguate Harvard and Chicago variants. Both have multiple variants (harvard-cite-them-right vs other Harvard flavours; chicago-author-date vs chicago-note-bibliography). Ask the user which one they want when they say "Harvard" or "Chicago" without specifying.
• WHO IRIS is a real differentiator. Most other citation tools cannot resolve WHO IRIS URLs. When the user shares a WHO publication, this skill is specifically the right tool.
• Batch single-tool calls, not loops. Sending five identifiers as one comma-separated batch returns five citations in one response; sending them as five separate tool calls multiplies round-trips and can hit rate limits.
• Pin the version when reproducibility matters. The MCP server calls the current production endpoint, which produces deterministic output for the same input + cache state. The x-scholar-cache header in the metadata block makes cache-hit vs cache-miss visible.

When NOT to Use This Skill

• Searching for papers by topic or keyword, or traversing citation networks — use a Semantic Scholar or OpenAlex MCP wrapper for literature discovery. Scholar Sidekick assumes you already have an identifier.
• Reading or editing a Zotero library — use zotero-mcp for stateful library access (search, annotate, manage collections).
• Repairing references inside a manuscript file (.tex, .bib, .md, .docx) — use citecheck for that workflow.
• No MCP host available — use the scholar-sidekick-api skill, which calls the same REST service over curl with no install and no key.

These tools compose well — Scholar Sidekick handles the formatting layer once another tool has produced the identifier.