PaperManager

MCP Server (Claude Desktop)

PaperManager ships with an MCP (Model Context Protocol) server at backend/mcp_server.py. This lets Claude Desktop interact with your library directly via natural language.

Setup

1 — Locate your Python binary

conda activate papermanager
which python
# e.g. /Users/you/miniconda3/envs/papermanager/bin/python

2 — Configure Claude Desktop

Edit ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "paperManager": {
      "command": "/path/to/conda/envs/papermanager/bin/python",
      "args": ["/path/to/PaperManager/backend/mcp_server.py"]
    }
  }
}

Replace the paths with your actual paths. Restart Claude Desktop.

3 — Verify

Ask Claude: “Can you list my recent papers?” — Claude should call search_papers and return results from your library.

Available Tools

Tool	Description
`search_papers`	Search by keyword, tag, topic, project, or person
`get_paper_detail`	Full paper metadata for one paper
`chat_with_paper`	Ask a question about a paper’s content
`add_note`	Write or update a paper’s markdown note
`get_note`	Read a paper’s note
`tag_paper_with`	Add a tag to a paper
`add_topic`	Link a research topic to a paper
`link_person_to_paper`	Link a person to a paper with a role
`add_paper_metadata`	Add a paper by metadata only (no PDF)
`list_projects`	List all projects
`list_project_papers`	Papers in a specific project
`add_to_project`	Add a paper to a project
`list_tags`	All tags with paper counts
`list_topics`	All topics with paper counts
`list_people`	All people
`get_person_papers`	Papers associated with a person
`add_person`	Create a person node
`create_project`	Create a new project

Example Interactions

You: add a note to the Attention is All You Need paper saying
     "@Jan is working on a follow-up to this approach"

Claude: [calls add_note] → note saved, MENTIONS relationship
        created in Neo4j for Jan

You: what papers do I have tagged "to-read" about protein folding?

Claude: [calls search_papers with tag="to-read", topic="Protein Folding"]
        → returns matching papers

You: create a project called "NeurIPS 2025 prep" and add all papers
     tagged "important" to it

Claude: [calls create_project, then search_papers + add_to_project
        for each paper]

Architecture

The MCP server and FastAPI backend share the same service and database layers:

flowchart LR
    CD["Claude Desktop"] -->|"MCP protocol"| MCP["mcp_server.py"]
    Browser["Browser"] -->|"HTTP"| FastAPI["FastAPI routers/"]
    MCP --> Tools["tools/*.py"]
    FastAPI --> Routers["routers/*.py"]
    Tools --> DB["db/queries/"]
    Routers --> DB
    Tools --> Services["services/"]
    Routers --> Services
    DB --> Neo4j["Neo4j Aura"]
    Services --> Drive["Google Drive"]
    Services --> AI["Claude / Ollama"]

!!! note “PDF upload is browser-only” File upload is intentionally not exposed as an MCP tool — it requires the browser drag-and-drop flow.

Running the MCP Server Standalone

The server can also be run without Claude Desktop for testing:

conda activate papermanager
cd /path/to/PaperManager
python backend/mcp_server.py

This starts the MCP server in stdio mode, ready to accept tool calls.