A traceable research workflow with Scholion MCP

A bounded question before a connected workflow

Suppose you want to start a reading list on platform governance, with a focus on how complementors manage dependence on platform owners. The goal of this worked example is to retrieve an inspectable set of papers and follow one citation trail—not ask an agent for a comprehensive review from a single prompt.

The tool names and argument shapes below were checked against Scholion’s MCP implementation. The transcript is illustrative, not an executed client session. No search result, credit balance or synthesized finding is presented as a live output. Actual results depend on catalog coverage and the date of execution.

Connect using a supported authentication pattern

Use a client that supports remote Streamable HTTP and static request headers. Scholion’s current setup page documents Claude Code and other configurable static-header clients. OAuth-only connector flows are not supported; a client that only offers “sign in with OAuth” cannot use this procedure.

Create an API key in your account, then keep it in the client’s protected configuration. This command is a template:

claude mcp add --transport http scholion https://scholion.ai/mcp \
  --header "Authorization: Bearer YOUR_SCHOLION_API_KEY"

The placeholder is deliberately not a working key. Never place a real key in a URL, a manuscript, a shared prompt, a screenshot or a committed configuration file. Be aware that shell history can retain commands; use your client’s secure configuration method when available. Scholion also accepts the X-API-Key header; send one authentication header, not both. Reload the client and enable the Scholion tools if they do not appear.

Step 1: ask for a limited search

Tell the host agent: “Search these terms, show bibliographic evidence, and ask before expanding the search or running additional paid calls.”

{
  "tool": "search_keyword",
  "arguments": {
    "query": "complementor | ecosystem",
    "journals": "MISQ,ISR,JMIS,JAIS",
    "year_from": 2020,
    "year_to": 2025,
    "limit": 10,
    "offset": 0,
    "sort": "year_desc"
  }
}

This is a readable description of a tool invocation for the host agent, not a raw JSON-RPC request to paste into an HTTP client. Client-specific tool prefixes may differ. This example searches for either “complementor” or “ecosystem.” In keyword search, spaces combine terms with AND and the vertical bar groups unquoted terms with OR. Quoted phrases are required phrases, not alternatives in an OR group; run a separate "platform governance" query if that phrase is your focus. The year bounds apply to the API’s publication-year semantics: issue year when present, otherwise online-publication year. That differs from an editorial snapshot explicitly restricted to issue year only.

Inspect returned titles, journals, years and DOIs before broadening the scope. A matching term can be incidental. Ask the agent to separate directly relevant governance studies from application-specific papers where a platform is merely the setting.

Step 2: resolve and inspect a known paper

A real catalog-verified starting point is the ISR paper on Android platform forks, DOI 10.1287/isre.2018.0786. It lies outside the recent-year search above, so treat it as a deliberate earlier addition rather than claiming it appeared in those results.

{
  "tool": "get_paper_by_doi",
  "arguments": { "doi": "10.1287/isre.2018.0786" }
}

Next, use get_paper_references or get_paper_related with that DOI and a small limit. References are outgoing links; related papers use shared-reference evidence, which is not a guarantee of identical research questions. Keep identifiers exactly as returned. Public APIs use DOIs or opaque paper identifiers, not internal database IDs.

Step 3: write a traceable note

Ask for a small table of paper title, DOI, why it belongs, evidence available and unresolved questions. If only an abstract was retrieved, the note must say “abstract-based,” not imply full-text reading. An appropriate illustrative note is: “Check whether the study’s account of boundary resources explains cooperation, competition, or both.” It is a question to examine, not a fabricated tool result.

Stop when the evidence is sufficient for the bounded question. If search returns little, broaden one dimension at a time and record the change. Semantic search can help discover vocabulary, but its index size is not the number of papers about your topic.

Credits, failures and reproducibility

MCP tool execution, REST calls and API calls made by downloadable skills share API credits, daily quotas and endpoint pricing. Connecting a client does not create a second free pool. Check current account/API documentation rather than relying on a fixed price in this article. Set a call budget with your agent and verify usage in your account.

For a 401, check the key and header configuration without pasting the key into chat. For quota or rate-limit responses, respect retry guidance; do not tell an agent to retry indefinitely. Save the query, parameters, retrieval date and selected DOI list without credentials. That small record is more reproducible than a polished answer with no retrieval history.