Scholar Sidekick declares its tools on navigator.modelContext. An AI agent running in your browser calls them by name - verify a citation, audit a bibliography, resolve an identifier - instead of scraping the page.
No install. No account. No API key. It self-disables in browsers that do not expose the API.
Experimental. WebMCP is an emerging standard and the browser API is still changing.
An agent that reads a citation page has to interpret rendered text, and it can misread it. WebMCP removes that step. The page hands the agent a typed tool it can call directly, and the answer comes back as structured data from the same endpoints that power every other Scholar Sidekick integration - so the agent gets the resolved record, not its own reading of a screen.
This is a transport, not a new product. The seven tools are the same ones the MCP server offers to agents outside the browser. WebMCP simply removes the install step for an agent that is already in the tab.
scholar-sidekick.com in a browser whose in-browser Model Context API is enabled.navigator.modelContext. If the API is absent it stops, and the page behaves exactly as before.verifyCitation - Checks a claimed citation against the record its identifier actually resolves to - the real-identifier-plus-invented-title pattern documented by Topaz et al. (Lancet, 2026).auditBibliography - Audits a whole bibliography (BibTeX, RIS, CSL-JSON, or a claims array) in one call: a verdict per entry plus a retraction check and a corpus summary. Capped at 25 entries.checkRetraction - Reports whether one work is retracted, corrected, or carries an expression of concern (Crossref / Retraction Watch).checkOpenAccess - Reports one work's open-access status and its best legal full-text URL (Unpaywall).resolveIdentifier - Turns a DOI, PMID, PMCID, ISBN, arXiv ID, ISSN, ADS bibcode, or WHO IRIS URL into structured CSL-JSON metadata.formatCitation - Formats identifiers into a finished citation: Vancouver, AMA, APA, IEEE, CSE, or any of 10,000+ CSL styles.exportCitation - Exports identifiers to a bibliography file: BibTeX, RIS, EndNote, RefWorks, NBIB, Zotero RDF, CSL-JSON, or CSV.All seven are read-only. None of them writes data, takes payment, or changes anything you own.
https://scholar-sidekick.com over HTTPS.WebMCP is described in the Chrome developer documentation and discussed in the W3C Web Machine Learning community group, which maintains a list of sites that implement it - Scholar Sidekick among them. The standard is young, so treat this surface as experimental: the tools are stable, the way a browser exposes them is not.
WebMCP lets a web page declare typed tools to an AI agent running in the browser, through the navigator.modelContext API. The agent calls a named tool with structured arguments instead of reading and clicking the page. It is an emerging standard, discussed in the W3C Web Machine Learning community group and shipping experimentally in Chrome.
Nothing. Open scholar-sidekick.com in a browser whose in-browser Model Context API is enabled, and the tools register themselves. There is no extension, no account, and no API key.
Seven: verifyCitation, auditBibliography, checkRetraction, checkOpenAccess, resolveIdentifier, formatCitation, and exportCitation. All seven are read-only - none of them writes anything or changes your data. Some third-party directories still list six, because auditBibliography joined the surface after they indexed the site.
Nothing at all. The registration script feature-detects navigator.modelContext and stops if the API is absent. No error, no console noise, and no effect on the page. Almost every visitor is in this case today.
An agent that scrapes a citation page reads rendered text and can misread it. A typed tool call returns the resolved record itself, with the same determinism as the REST API: the same input gives the same verdict, and the verdict is computed in code against the upstream registries, not inferred by a language model.
The tools are the same seven; the transport differs. The MCP server is for agents outside the browser - Claude, Cursor, ChatGPT - over stdio or Streamable HTTP. WebMCP is for an agent already inside the browser tab, with no install step at all.
The underlying API is experimental and the standard is still moving, so treat the surface as experimental too. The tools themselves are thin clients over the public Scholar Sidekick API, which is stable and versioned. If the browser API changes shape, the registration changes; the endpoints behind it do not.