Paper Search MCP

Pick a task and get started

Obtain the pinned source and inspect its tools, then connect that same checkout to your client. The six tasks cover discovery, citation identity, retrieval, citation expansion, section locations, and existing institutional access.

1. Get the reviewed source and list its tools

   Copy install command

   Run from a location without an existing paper-search-mcp subdirectory. The final line lists this checkout’s registered tools rather than querying papers; the first uv run installs dependencies. The checkout is pinned to the reviewed commit, which the matching PyPI version label 0.1.4 cannot substitute for.

   git clone https://github.com/openags/paper-search-mcp.git
   cd paper-search-mcp
   git checkout cab1b5cc276a0cd004a0044003926a6440088049
   mkdir -p downloads
   uv run -m paper_search_mcp.cli tool --list

   Read the author’s setup guide

   Connect the same checkout to an MCP client

   Copy install command

   Claude Desktop uses ~/Library/Application Support/Claude/claude_desktop_config.json on macOS and %APPDATA%\Claude\claude_desktop_config.json on Windows. Merge this entry into mcpServers, preserving existing services; other hosts have their own MCP configuration entry point. Replace both /absolute/path/to/... values with the actual checkout path, and use an absolute uv executable path if the client cannot find it. Save, restart the client, and confirm tool discovery before sending a task. The download directory is also the section tool’s readable root.

   {
     "mcpServers": {
       "paper-search-mcp": {
         "command": "uv",
         "args": [
           "run",
           "--directory",
           "/absolute/path/to/paper-search-mcp",
           "-m",
           "paper_search_mcp.server"
         ],
         "env": {
           "PAPER_SEARCH_MCP_TRANSPORT": "stdio",
           "PAPER_SEARCH_MCP_SECTION_PDF_ROOT": "/absolute/path/to/paper-search-mcp/downloads"
         }
       }
     }
   }

   Read the author’s source-launch instructions

2. Give your AI this first task

   Copy prompt

   These editorial prompts are examples, not observed results. Replace the bracketed material and send one to an assistant with the tools connected. Tables and notes are prepared by the assistant; the service returns records, text, or file paths.

   Use the connected Paper Search MCP for this question: [one specific research question]. Keywords and synonyms: [search terms]. Include: [years, population or subject, and publication types]. Exclude: [out-of-scope material]. Confirm that search_papers is available before searching. Start with search_papers(query="[terms]", sources="arxiv,crossref,openalex", max_results_per_source=5). The limit is 5 per source. Do not use the CLI presets fast/fastest as MCP source names, or claim that year filters these three sources. Keep the actual arguments and raw response. Check sources_used, source_results, and errors, then screen by the real publication date and available abstract. Flag placeholder dates, missing abstracts, and source failures; an empty result is not proof that no relevant research exists. For a strict date range, use the dedicated Crossref/OpenAlex tools with a filter and record it. Build a candidate table with title, authors, year and date precision, source, paper_id, DOI, original URL, abstract availability, and an include/exclude/unresolved reason. Check duplicate titles and preprint versus published versions; automatic deduplication does not verify versions. State which sources this round covered and which remain unresolved. Do not infer paper findings from titles alone.

3. Check what you receive

   • A candidate table, actual query arguments, per-source counts, and failure notes.
   • A DOI or source ID, original link, and screening reason for each candidate; missing fields remain unresolved.
   • Checked dates, versions, and coverage. This round does not establish an exhaustive systematic-review search.
   Read the relevant tool documentation