crossref-mcp

English | 中文

An MCP server wrapping the Crossref REST API, exposing scholarly-metadata lookups (works, members, journals, funders, types, licenses, prefixes) as tools an LLM client can call.

Not affiliated with or endorsed by Crossref. Data is served live from the public Crossref API. Set CROSSREF_MAILTO to join Crossref's polite pool.

Features

• All major Crossref resource endpoints as 18 read-only tools.
• Dual transport: stdio (local) and Streamable HTTP (container).
• Polite pool (User-Agent + mailto), optional Crossref Plus token.
• Token-bucket rate limiting (auto-tuned from X-Rate-Limit-* headers) with exponential backoff honoring Retry-After.
• Cursor deep paging, field select, raw vs simplified output.
• Optional X-API-Key HTTP auth (/health exempt), /health endpoint.
• Optional Redis response cache + cross-replica rate limiting.
• Optional Caddy TLS reverse proxy for public deployment.

Quick start (local, stdio)

uv sync
export CROSSREF_MAILTO="you@example.com"   # recommended (polite pool)
uv run crossref-mcp                         # starts a stdio MCP server

Inspect with the MCP Inspector: uv run mcp dev src/crossref_mcp/server.py.

Configuration

Tools

Works — search_works, get_work (by DOI), get_work_references (capped at 50), get_work_quality (registration agency). Members / Journals / Funders — search_members · get_member · get_member_works; search_journals · get_journal (ISSN) · get_journal_works; search_funders · get_funder · get_funder_works. Types / Licenses / Prefixes — list_types · get_type; list_licenses; get_prefix. Plus ping.

Shared parameters. Search/list tools take query (and query_bibliographic / query_author / query_title on works), filter (Crossref key:value,… syntax), sort + order, rows (≤1000) + offset, and select. Deep paging: set cursor="*" then reuse the returned next_cursor. All tools return simplified fields by default; pass raw=true for the full Crossref JSON.

Connecting an MCP client

stdio (Claude Desktop / Cursor) — claude_desktop_config.json:

{
  "mcpServers": {
    "crossref": {
      "command": "uvx",
      "args": ["crossref-mcp"],
      "env": { "CROSSREF_MAILTO": "you@example.com" }
    }
  }
}

Or run the container over stdio: "command": "docker", "args": ["run", "-i", "--rm", "-e", "CROSSREF_MAILTO", "heyinnaneo/crossref-mcp"].

Streamable HTTP — once the container is up, point an HTTP-capable client at http://localhost:8000/mcp. If MCP_API_KEY is set, send it as X-API-Key.

Use as a skill (no server)

Prefer not to run a server? skill.md is a self-contained skill that teaches an agent to call the Crossref REST API directly — it only needs an HTTP/fetch capability (no install, no key, no container). Trade-off: the agent issues raw HTTP and parses Crossref JSON itself, without the MCP server's polite-pool throttling, field simplification, or structured errors.

Claude Code (Agent Skill):

mkdir -p ~/.claude/skills/crossref-api
curl -fsSL https://raw.githubusercontent.com/jtl-neo/crossref-api-MCP/main/skill.md \
  -o ~/.claude/skills/crossref-api/SKILL.md

(Project-scoped instead: place it at .claude/skills/crossref-api/SKILL.md in your repo.)

Other agents (Cursor, Claude Projects, custom): paste the contents of skill.md into your system prompt / project knowledge / rules file.

Docker / HTTP deployment

cp .env.example .env          # set CROSSREF_MAILTO
docker compose up -d          # pulls heyinnaneo/crossref-mcp, HTTP on :8000
curl http://localhost:8000/health

Serves Streamable HTTP at /mcp and /health (compose health check). Set MCP_API_KEY to require X-API-Key on /mcp. To build locally, uncomment build: . in docker-compose.yml.

Public deployment (TLS / reverse proxy) — optional

Put it behind the bundled Caddy proxy (Caddyfile + docker-compose.proxy.yml): TLS termination (auto Let's Encrypt), HTTP→HTTPS, SSE streaming, backend not published on the host.

cp .env.example .env   # add DOMAIN=, ACME_EMAIL=, MCP_API_KEY=
docker compose -f docker-compose.proxy.yml up -d
curl https://your-domain/health

Defense in depth: proxy edge (optional IP allowlist / Basic auth) + MCP_API_KEY app layer; both leave /health open. Needs a domain with DNS pointing at the host and ports 80/443 reachable. Caddy rate limiting needs the caddy-ratelimit plugin (custom build).

Redis cache + cross-replica rate limiting — optional

Set REDIS_URL for a response cache (cache-aside on the raw envelope; mailto/ secrets excluded from keys) and, with RATELIMIT_BACKEND=redis, a shared token bucket so replicas stay within Crossref's polite-pool rate.

docker compose -f docker-compose.redis.yml up -d
curl http://localhost:8000/health   # cache_enabled + ratelimit_backend + redis: up

Redis failures degrade gracefully (no cache, in-memory limiting); /health stays 200 with redis: down.

CI / publishing

.github/workflows/ci.yml: test (ruff + format + pytest with coverage on every push/PR), security (Trivy fs scan), build-push (multi-arch image to heyinnaneo/crossref-mcp on main / v* tags), and on tags publish-pypi (OIDC Trusted Publisher) + publish-registry (MCP registry).

Repository secrets (Settings → Secrets and variables → Actions): DOCKERHUB_USERNAME = heyinnaneo, DOCKERHUB_TOKEN = a Docker Hub Access Token (Read/Write). Tag a release (git tag v0.1.0 && git push origin v0.1.0); the tag must match pyproject.toml.

Install from the MCP registry

Published as io.github.jtl-neo/crossref-mcp (see server.json), with a PyPI package (uvx crossref-mcp) and an OCI image. Publishing is automated on version tags (publish-pypi + publish-registry); requires a public repo and a configured PyPI Trusted Publisher.

Security & trust

• Read-only. Every tool is a lookup, annotated readOnlyHint.
• Rate limiting. Single-process in-memory by default — don't run as a public proxy for heavy traffic; for replicas use RATELIMIT_BACKEND=redis.
• Polite pool. Use your own CROSSREF_MAILTO.
• Public HTTP. Set MCP_API_KEY and front with TLS.
• No bundled secrets. .env is excluded from image and git.

License

MIT. Bibliographic data comes from the public Crossref REST API; not affiliated with Crossref.

中文

English | 中文

把 Crossref REST API 封裝成 MCP server，將學術文獻 metadata 查詢（works、members、journals、funders、types、 licenses、prefixes）包成 LLM client 可呼叫的工具。

非 Crossref 官方、未經其背書。資料即時取自公開的 Crossref API。請設定 CROSSREF_MAILTO 以加入 Crossref 的 polite pool。

特色

• 主要 Crossref 資源端點，共 18 個唯讀工具。
• 雙傳輸：stdio（本機）與 Streamable HTTP（容器）。
• Polite pool（User-Agent + mailto）、可選 Crossref Plus token。
• Token-bucket 速率控制（依 X-Rate-Limit-* header 自動調整），429 指數退避並 尊重 Retry-After。
• cursor 深分頁、欄位 select、raw 與精簡輸出切換。
• 可選 X-API-Key HTTP 認證（/health 豁免）、/health 端點。
• 可選 Redis 回應快取 + 跨副本速率控制。
• 可選 Caddy TLS 反向代理，供公開部署。

快速開始（本機 stdio）

uv sync
export CROSSREF_MAILTO="you@example.com"   # 建議（polite pool）
uv run crossref-mcp                         # 啟動 stdio MCP server

用 MCP Inspector 檢視：uv run mcp dev src/crossref_mcp/server.py。

設定

工具