paper-tools-mcp 0.1.1

Standalone MCP server: deterministic paper identity resolution + PDF retrieval

paper-tools (MCP)

A standalone MCP server that turns "find a paper" into a deterministic capability. Two tools:

• resolve_paper(query) — any input (DOI, arXiv id/URL, SSRN, a bare title, a messy citation string) → decide whether it points to one real paper and return that paper's unique identifier (DOI, else OpenAlex Work ID), or ambiguous / not_found. Never guesses a binding.
• get_pdf(ids) — that identifier → actually downloads the open-access PDF, trying arXiv → Unpaywall → OpenAlex OA/green/SSRN. Returns the saved path, or an honest downloadable=false with human follow-up leads (SSRN, OA Button) when no open-access copy exists.

Built by an unattended Ralph loop from GOAL.md. Status and remaining work live in PROGRESS.md. This project is fully independent — it has no dependency on any other repository.

Setup

conda env create -f environment.yml
conda activate paper-tools-mcp
pytest -q            # offline golden-set suite
pytest -q -m live    # live smoke (hits real APIs)

Institutional access (your own subscription)

When a paper has no open-access copy, get_pdf tries the caller's own subscription (disable with institutional=False). Two legitimate mechanisms:

• Elsevier full-text API (covers 10.1016/...): export your own key — export ELSERVIER_API_KEY=... (optional ELSERVIER_INSTTOKEN=...). Uses the Elsevier Article Retrieval API, entitled via your institutional network/token.

• Publisher fetch (curl_cffi): follows the DOI to the publisher and grabs the PDF; works for publishers that grant access by IP when run on campus/VPN.

• Your Chrome login cookies (curl_cffi): the publisher fetch reuses your existing Chrome session via browser_cookie3 — stay logged into the publishers in Chrome. Works for publishers without a JS bot-wall. Disable with PAPER_TOOLS_NO_BROWSER_COOKIES=1.

• Strong layer (ON by default): one-click institutional fetch for publishers behind a Cloudflare human-check + SSO (SAGE → POMS/JMR/JM, OUP → JCR, INFORMS). Runs whenever a desktop display is present (chromium auto-installs on first use); opt out with PAPER_TOOLS_NONINTERACTIVE=1 (e.g. on headless servers, where it degrades to leads). When the automated cascade can't get a gated paper, get_pdf pops ONE browser window — already seeded with your Chrome login cookies — you clear the one-time "are you human" check (and sign in if needed) once per publisher, and the tool downloads that paper and reuses the session for every later paper from the same publisher. The human check is solved by you (never auto-defeated); everything else is automatic.

Install = environment solved (no manual setup)

Installing the package pulls all Python dependencies; the chromium browser for the strong layer auto-installs on first use — no separate conda/pip/ patchright install steps.

Zero-setup registration (uv creates an isolated env and installs on first run):

// .mcp.json
{
  "mcpServers": {
    "paper-tools": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/<you>/paper-tools-mcp", "paper-tools-mcp"],
      "env": { "PAPER_TOOLS_INTERACTIVE": "1", "ELSERVIER_API_KEY": "<optional>" }
    }
  }
}

Or with pip: pip install git+https://github.com/<you>/paper-tools-mcp then use "command": "paper-tools-mcp". (A desktop display is required only for the strong interactive layer; the OA/API layers run anywhere.)

Cross-platform: Linux, macOS, and Windows. The OA/API layers run anywhere. The strong interactive layer uses your local Chrome cookies (browser_cookie3 decrypts per-OS: Linux keyring / macOS Keychain / Windows DPAPI) and pops a desktop window, so it needs a logged-in Chrome + a desktop session. Launch flags are auto-selected per OS (--no-sandbox only on Linux). Note: very recent Chrome on Windows uses app-bound cookie encryption that browser_cookie3 may not read — sign in once in the popped window if cookie reuse misses.

Register with Claude Code (dev / conda)

// .mcp.json / Claude Code MCP config
{
  "mcpServers": {
    "paper-tools": {
      "command": "conda",
      "args": ["run", "-n", "paper-tools-mcp", "python", "-m", "paper_tools.server"]
    }
  }
}