scholar-mcp

A FastMCP server for the scholarly citation landscape (papers, patents, books, and standards), giving LLMs a unified way to search, cross-reference, and retrieve prior art across all four source types via Semantic Scholar, EPO Open Patent Services, Open Library, and standards bodies (NIST, IETF, W3C, ETSI), with OpenAlex enrichment and optional docling-serve PDF/full-text conversion.

Documentation | Config wizard | PyPI | Docker

Features

Source domains

• Papers: full-text search with year/venue/field/citation filters; single-paper lookup by DOI, S2 ID, arXiv ID, ACM ID, or PubMed ID; author profile and name search; forward citations, backward references, BFS graph traversal, shortest-path bridge discovery; recommendations from positive/negative examples; BibTeX/CSL-JSON/RIS citation generation with OpenAlex venue enrichment.
• Patents: search across 100+ patent offices via EPO OPS with CPC/applicant/inventor/jurisdiction filters; bibliographic, claims, description, family, legal, and citations sections; NPL-to-paper resolution via Semantic Scholar and paper-to-patent citation discovery. EPO credentials are optional; other domains work without them.
• Books: Open Library search by title/author/keywords, no API key required; lookup by ISBN-10/13 or by Open Library work/edition ID; subject-based recommendations sorted by popularity; Google Books excerpts and preview links; WorldCat permalinks for library discovery; cover image caching. Papers with an ISBN in externalIds are automatically enriched with publisher, edition, cover URL, and subject data from Open Library.
• Standards: identifier resolution, search, and metadata retrieval for NIST, IETF, W3C, and ETSI standards, with optional full-text fetch and Markdown conversion via docling. Tier 2 ISO, IEC, IEEE, Common Criteria (CC), and CEN/CENELEC metadata (including ISO/IEC/IEEE joint standards and the CC ↔ ISO/IEC 15408 cross-link) is synced locally via sync-standards. ISO, IEC, IEEE have a live-fetch fallback for unsynced identifiers; CC and CEN have no live API and require a sync first. Citations matching standards patterns (RFC, ISO, NIST SP, IEEE, EN, CC) are automatically enriched with structured standard_metadata including identifier, title, body, status, and full-text URL when available (see docs/guides/standards.md).

Cross-cutting

• Enrichment pipeline: phased enrichment from multiple sources: OpenAlex (OA status, affiliations, funders, concepts), CrossRef (publisher, page ranges, container titles), Google Books (preview links, excerpts), and Open Library (book metadata). Runs automatically on paper and book results.
• PDF conversion: download open-access PDFs and convert to Markdown via docling-serve, with optional VLM enrichment for formulas and figures; automatic fallback to ArXiv, PubMed Central, and Unpaywall when Semantic Scholar has no OA link; direct URL download for PDFs found elsewhere.
• Intelligent caching: SQLite-backed cache with per-table TTLs (30 days for papers/authors, 7 days for citations/references) and identifier aliasing.
• Authentication: bearer token, OIDC (OAuth 2.1), or both simultaneously (multi-auth).
• Multi-transport: stdio (Claude Desktop), HTTP (streamable-http), and SSE transports.
• Linux packages: .deb and .rpm packages with systemd service and security hardening.

Coverage by domain

Per-domain depth is uneven. Papers currently have the richest tool surface (citation graph, recommendations, cross-referencing to all three other domains); standards are the leanest. That reflects public data availability, not a value hierarchy: writing a paper typically needs all four source types for citations and prior art. Parity work is tracked in GitHub issues and milestones; the roadmap shows intent, not a completeness commitment.

What you can do with it

With this server mounted in an MCP client (Claude, etc.), you can:

• Survey a field: "Find the 20 most-cited papers on graph neural networks from 2020 to 2024 and draft a literature review outline." Composes search_papers + get_citations + enrich_paper.
• Trace a citation path: "What's the shortest citation path from 'Attention is All You Need' to 'RLHF for dialogue agents'?" Uses find_bridge_papers + get_citation_graph.
• Cross-reference prior art: "For this patent family, list academic papers it cites and any books or standards that show up in the description." Composes get_patent + batch_resolve + standards/book enrichment.
• Generate a bibliography: "Emit BibTeX for these 30 DOIs with OpenAlex venue data." Uses generate_citations.
• Look up a standard: "What's the latest status of RFC 9000, and fetch the Markdown full text." Uses resolve_standard_identifier + get_standard.

Installation

From PyPI

pip install pvliesdonk-scholar-mcp

If you add optional extras via the PROJECT-EXTRAS-START / PROJECT-EXTRAS-END sentinels in pyproject.toml, document them below:

Scholar-mcp ships two optional-dependency groups:

• [mcp]: installs FastMCP; required to run scholar-mcp serve and expose tools over stdio/HTTP.
• [all]: currently identical to [mcp]; reserved for future optional backends.

For MCP-server usage:

pip install 'pvliesdonk-scholar-mcp[mcp]'
# or, without installing into the environment:
uvx --from pvliesdonk-scholar-mcp scholar-mcp serve

Installing the bare pvliesdonk-scholar-mcp package is enough for library use (from scholar_mcp import ...) but the scholar-mcp serve CLI requires [mcp].

From source

git clone https://github.com/pvliesdonk/scholar-mcp.git
cd scholar-mcp
uv sync --all-extras --all-groups

Docker

docker pull ghcr.io/pvliesdonk/scholar-mcp:latest

To run the newest merged code instead of the newest release, use the rolling edge tag. It is rebuilt on every merge to main and carries no version identity. See Image tags for the full tag list.

docker pull ghcr.io/pvliesdonk/scholar-mcp:edge

A compose.yml ships at the repo root as a starting point. Copy .env.example to .env, edit, and docker compose up -d.

To attach a remote Python debugger (development only; the protocol is unauthenticated), see Remote debugging.

Linux packages (.deb / .rpm)

Download .deb or .rpm packages from the GitHub Releases page. Both install a hardened systemd unit; env configuration is sourced from /etc/scholar-mcp/env (copy from the shipped /etc/scholar-mcp/env.example).

Claude Desktop (.mcpb bundle)

Download the .mcpb bundle from the GitHub Releases page and double-click to install, or run:

mcpb install scholar-mcp-<version>.mcpb

Claude Desktop prompts for required env vars via a GUI wizard, with no manual JSON editing needed.

For manual Claude Desktop configuration and setup options, see Claude Desktop deployment.

Release channels

Artifacts ship on three channels. Each row lists exactly what that channel publishes.

Pre-releases reach PyPI so that a candidate's .mcpb bundle installs: the bundle points at PyPI rather than carrying the code. Ordinary installers never see them, because a PEP 440 resolver skips pre-releases unless the requirement pins one or you pass --pre. Ask for a candidate by name with pip install pvliesdonk-scholar-mcp==X.Y.ZrcN. PyPI spells it in the PEP 440 canonical form, while tags use SemVer. Rolling pointers are ordering-aware, so a patch release cut from an old release/X.Y branch never moves latest-style tags back to older content, and a candidate for an already-released version never moves rc. See Release process for the full model.

Quick start

scholar-mcp serve                                # stdio transport
scholar-mcp serve --transport http --port 8000   # streamable HTTP

For library usage (embedding the domain logic without the MCP transport), import from the scholar_mcp package directly. Backend clients live under src/scholar_mcp/_s2_client.py, _epo_client.py, _openlibrary_client.py, and _standards_client.py.

Server info

The server registers a built-in get_server_info tool (via fastmcp_pvl_core.register_server_info_tool) so operators can confirm the deployed version with a single MCP call. The default response carries server_name, server_version, and core_version. Servers that talk to a remote upstream wire upstream version reporting inside the DOMAIN-UPSTREAM-START / DOMAIN-UPSTREAM-END sentinel in src/scholar_mcp/server.py; see CLAUDE.md for the wiring pattern.

Configuration

Core environment variables shared across all fastmcp-pvl-core-based services:

Domain-specific variables go below under Domain configuration.

Authorization (opt-in)

This server inherits opt-in per-subject authorization from fastmcp-pvl-core. The default posture is off: every authenticated caller can use every tool, resource, and prompt. Turn it on by pointing SCHOLAR_MCP_ACL_PATH at a TOML ACL file; the middleware is installed only when the path is set, and individual tools opt in by declaring meta={"required_scope": "<scope>"} at registration. A tool without required_scope is unrestricted regardless of caller.

Wire it in by uncommenting the acl_path field in src/scholar_mcp/config.py and the AuthorizationMiddleware stanza in src/scholar_mcp/server.py; both ship as commented stubs in the scaffold.

ACL TOML schema

[subjects]
"user:alice@example.com" = ["read", "write"]
"user:admin@example.com" = ["*"]              # wildcard — any required scope passes
"service:ci-bot"         = ["read"]
"local"                  = ["*"]              # auth-disabled subject (no bearer / OIDC vars set)

• Subject strings are opaque. The <kind>:<id> convention is documentation only; the library treats each subject as a literal string.
• * is the only library-treated special scope: it grants every required scope. Subject-side wildcards (* as an ACL key) are rejected at load time.
• Scope vocabulary is domain-defined. Per-project or per-folder gating is encoded into the scope string itself, such as read:project-foo or write:vault/personal; fastmcp-pvl-core treats every scope except * as opaque.

Subject ↔ bearer-token alignment

The subject string used as a value in the bearer-tokens TOML (SCHOLAR_MCP_BEARER_TOKENS_FILE) is the same string used as a key in the ACL TOML. Same string, opposite roles, so keep the two files consistent when adding or removing a principal. See Mapped bearer tokens in the authentication guide for the bearer-tokens TOML schema.

In single-token mode (SCHOLAR_MCP_BEARER_TOKEN) every authenticated caller shares one subject, the library's default (currently "bearer-anon"); override it with SCHOLAR_MCP_BEARER_DEFAULT_SUBJECT; reference that string as the ACL key. When no auth is configured (no SCHOLAR_MCP_BEARER_TOKEN, SCHOLAR_MCP_BEARER_TOKENS_FILE, or OIDC env vars set, which is common in stdio dev rigs but also possible on HTTP), every request resolves to the literal subject "local". Reference that string as the ACL key for un-authenticated local sessions.

Authentication

Callers authenticate via a bearer token or OIDC (mutually exclusive). See the Authentication guide for setup, mapped multi-subject tokens, OIDC, and troubleshooting.

Post-scaffold checklist

After copier copy and gh repo create --push:

1. Fill in the DOMAIN blocks (every section marked with a DOMAIN sentinel comment) in this README and in CLAUDE.md. The GENERATED-ENV-TABLE-* regions are not DOMAIN blocks; the config generator owns them and rewrites them on every run.
2. Configure GitHub secrets (see below).
3. Install dev + docs tooling: uv sync --all-extras --all-groups.
4. Install pre-commit hooks: uv run pre-commit install.
5. Run the gate locally: uv run pytest -x -q && uv run ruff check --fix . && uv run ruff format . && uv run mypy src/ tests/.
6. Push the first commit. CI should be green.

GitHub secrets

CI workflows reference three repository secrets. Configure them via Settings → Secrets and variables → Actions or with gh secret set:

gh secret set RELEASE_TOKEN
gh secret set CODECOV_TOKEN
gh secret set CLAUDE_CODE_OAUTH_TOKEN

Dependency updates are handled by Renovate (renovate.yml), which reuses RELEASE_TOKEN. It maintains uv.lock and auto-merges patch/minor bumps once the CI Success check is green; bootstrap.yml enables auto-merge and applies the repository rulesets (.github/rulesets/) on first push. See Repository Protection for the per-branch posture and bypass model. GitHub Actions are updated in the copier template and arrive via copier update, not per-repo.

GITHUB_TOKEN is auto-provided; no action needed.

Local development

The PR gate (matches CI):

uv run pytest -x -q                                  # tests
uv run ruff check --fix . && uv run ruff format .    # lint + format
uv run mypy src/ tests/                              # type-check

Pre-commit runs a subset of the gate on each commit; see .pre-commit-config.yaml for details, or CLAUDE.md for the full Hard PR Acceptance Gates.

Troubleshooting

Moving a scaffolded project

uv sync creates .venv/bin/* scripts with absolute shebangs pointing at the venv Python. If you move the repo after scaffolding (mv /old/path /new/path), uv run pytest fails with ModuleNotFoundError: No module named 'fastmcp' because the stale shebang resolves to a different interpreter than the venv's site-packages.

Fix:

rm -rf .venv
uv sync --all-extras --all-groups

uv run python -m pytest also works as a one-shot workaround (bypasses the stale entry-script shim).

uv.lock refresh after copier update

When copier update introduces new dependencies (such as a new extra added to pyproject.toml.jinja), the CI install step runs uv sync --locked, which fails against a stale lockfile. Run uv lock locally and commit the refreshed uv.lock alongside accepting the copier-update PR.

CI installs with --locked (and the review workflow with --frozen) so no job ever rewrites uv.lock in its own workspace: a job that re-locks hides the drift it just repaired, and a dirty workspace breaks any later git checkout in the same job. Lockfile drift then shows up as a red install step with a clear message, not as a silent mutation.

Links

• Documentation
• llms.txt
• FastMCP
• fastmcp-pvl-core

Domain configuration

Domain environment variables use the SCHOLAR_MCP_ prefix:

Domain-config fields are composed inside src/scholar_mcp/config.py between the CONFIG-FIELDS-START / CONFIG-FIELDS-END sentinels; env reads go through fastmcp_pvl_core.env(_ENV_PREFIX, "SUFFIX", default) so naming stays consistent, and field invariants go in __post_init__ between the CONFIG-VALIDATE-START / CONFIG-VALIDATE-END sentinels. Each field's metadata help and tags generate the table above directly, so keep them accurate and complete.

Scholar-mcp pings Semantic Scholar once on startup and every 7 days thereafter to keep the configured key from being removed for inactivity (Semantic Scholar may remove keys unused for 60+ days). If S2 starts rejecting the key with 403 Forbidden, this shows up in the server logs as s2_key_forbidden (on real tool calls) or s2_keepalive_key_forbidden (from the background keepalive); grep for either to confirm a dead key versus a transient upstream issue.

Key design decisions

• Library-first, MCP-optional. The core domain logic (S2/EPO/Open Library/standards clients, enrichment pipeline, cache) is importable without FastMCP; the MCP server is a thin async wrapper. Enables reuse in scripts, notebooks, and other servers.
• Sync domain code, async MCP layer. Backend clients are synchronous; MCP tools call them via asyncio.to_thread(). Simpler client code, explicit offloading at the transport boundary.
• SQLite cache with per-table TTLs and identifier aliases. Papers / authors last 30 days, citations / references 7 days. DOI ↔ S2 ID ↔ arXiv ID aliasing survives across cache clears so repeated enrichment hits the same row.
• Read-only by default. Write-tagged tools (PDF download/convert, patent PDF) are hidden unless SCHOLAR_MCP_READ_ONLY=false. Safer default for first-run.
• Slow work becomes a background job. Every tool whose work can run long runs through the fastmcp-pvl-core jobs layer, whether the slow part is a docling conversion, an EPO throttle being waited out, or a graph walk making one request per node. A call that beats SCHOLAR_MCP_JOBS_SOFT_DEADLINE_S returns its result directly; a slower one returns a handle to poll with get_job_result. No tool decides in advance whether to go background, so a cache hit needs no special case.
• EPO throttling is waited out, not queued. The traffic light is consulted before every request and cached for a minute, so a retry sooner than that would re-read the cache rather than ask again. Each backoff outlasts the cache; an exhausted daily quota is reported immediately instead, since it will not clear today.
• Tier 2 standards sync out-of-band. ISO/IEC/IEEE/CC/CEN catalogues come from community Relaton dumps via scholar-mcp sync-standards, not live at runtime, which avoids paywalled-HTML scraping and keeps tool calls fast.

Quick Start details

stdio transport (Claude Desktop / MCP clients)

uvx --from pvliesdonk-scholar-mcp scholar-mcp serve

API key optional but recommended: The server works without a Semantic Scholar API key, but unauthenticated requests are limited to ~1 req/s and will hit 429 throttles quickly during multi-step operations like citation graph traversal. Request a free key to get ~10 req/s.

Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "scholar": {
      "command": "uvx",
      "args": ["--from", "pvliesdonk-scholar-mcp", "scholar-mcp", "serve"],
      "env": {
        "SCHOLAR_MCP_S2_API_KEY": "your-key"
      }
    }
  }
}

HTTP transport

uvx --from pvliesdonk-scholar-mcp scholar-mcp serve --transport http --port 8000

Claude Code plugin

/plugin marketplace add pvliesdonk/claude-plugins
/plugin install scholar-mcp@pvliesdonk

Syncing Tier 2 standards catalogues

Tier 2 bodies (ISO, IEC, IEEE, CC, CEN) are populated from community-curated bulk dumps rather than live-scraped at MCP-server runtime. Run the sync on first install and periodically thereafter:

scholar-mcp sync-standards            # all registered bodies
scholar-mcp sync-standards --body ISO # only ISO
scholar-mcp sync-standards --body IEEE # only IEEE
scholar-mcp sync-standards --body CC   # only Common Criteria
scholar-mcp sync-standards --body CEN # only CEN/CENELEC
scholar-mcp sync-standards --force    # re-sync even if upstream SHA is unchanged

Schedule via cron, launchd, or a systemd timer. Weekly is sufficient; standards change slowly. First sync can take several minutes; subsequent runs that find no upstream changes exit within seconds.

MCP Tools

29 tools, organised by scholarly source type.

Papers

Search & retrieval

Citation graph

Recommendations & citation generation

Patents

Patent tools are hidden when SCHOLAR_MCP_EPO_CONSUMER_KEY and SCHOLAR_MCP_EPO_CONSUMER_SECRET are not set. fetch_patent_pdf is also write-tagged and hidden when SCHOLAR_MCP_READ_ONLY=true.

Books

Papers with an ISBN in their externalIds are automatically enriched with book_metadata (publisher, edition, cover URL, subjects, and more) from Open Library when fetched via get_paper, get_citations, get_references, or get_citation_graph. Book records also include worldcat_url (when ISBN-13 is present), google_books_url, and snippet from Google Books enrichment. Cover images can be downloaded and cached locally via get_book.

Standards

Tier-1 bodies (NIST, IETF, W3C, ETSI) are supported with full metadata and optional full-text conversion. Tier-2 bodies (ISO, IEC, IEEE, CC, CEN/CENELEC) are populated locally via scholar-mcp sync-standards.

Cross-source Utility

PDF Conversion (requires docling-serve)

PDF tools are write-tagged and hidden when SCHOLAR_MCP_READ_ONLY=true (the default). fetch_patent_pdf (above) and the get_standard full-text mode cover the patent and standards equivalents.

Job Polling

Tools answer directly when the work is quick, including on a cache hit. A slower call returns {"status": "working", "job_id": "...", "poll_with": "get_job_result"}; poll with the tool the handle names until the status is terminal.

Docker Compose

services:
  scholar-mcp:
    image: ghcr.io/pvliesdonk/scholar-mcp:latest
    restart: unless-stopped
    environment:
      SCHOLAR_MCP_S2_API_KEY: "${SCHOLAR_MCP_S2_API_KEY}"
      SCHOLAR_MCP_DOCLING_URL: "http://docling-serve:5001"
      SCHOLAR_MCP_VLM_API_URL: "${VLM_API_URL:-}"
      SCHOLAR_MCP_VLM_API_KEY: "${VLM_API_KEY:-}"
      SCHOLAR_MCP_CACHE_DIR: "/data/scholar-mcp"
      SCHOLAR_MCP_READ_ONLY: "false"
    volumes:
      - scholar-mcp-data:/data/scholar-mcp
    labels:
      - "traefik.enable=true"
      - "traefik.http.routers.scholar-mcp.rule=Host(`scholar-mcp.yourdomain.com`)"

  docling-serve:
    image: ghcr.io/ds4sd/docling-serve:latest
    restart: unless-stopped

volumes:
  scholar-mcp-data:

Cache Management

# Show cache statistics (row counts, database size)
scholar-mcp cache stats

# Clear all cached data (preserves identifier aliases)
scholar-mcp cache clear

# Remove entries older than 30 days
scholar-mcp cache clear --older-than 30

# Override cache directory
scholar-mcp cache stats --cache-dir /path/to/cache