CiteNexus

CiteNexus MCP

Scholarly discovery and citations with evidence you can trace.

CiteNexus connects research assistants to open scholarly metadata, specialist academic indexes, and optional commercial databases through the Model Context Protocol. Search without an API key, resolve stable identifiers, check references against their sources, and export citations without asking a language model to invent the missing fields.

This is a citation and metadata service for the assistant you already use. It works alongside reference managers such as Zotero and writing environments that support MCP.

CiteNexus 0.2.3 includes the MCP service and the cite-nexus-wtfp companion. The WTF-P integration targets its separate 0.7.0 release. Marketplace listings and native client plugin qualification are tracked independently.

Quick start · WTF-P integration · Provider guide · Troubleshooting · Release information

• Useful without credentials: Crossref, DataCite and Europe PMC are the default search sources. arXiv and Semantic Scholar are also available without mandatory keys.
• Optional specialist and commercial access: OpenAlex, Google Scholar through SerpAPI or SearchApi, Scopus and Web of Science Starter have explicit configuration and provider selection.
• Source-backed records: namespaced identifiers, retrieval timestamps, field attribution, incomplete-author warnings and provider-specific citation counts.
• Deterministic export: BibTeX, RIS and CSL-JSON from retrieved metadata. No model is required.
• Current MCP: Python SDK 2.2, the 2026-07-28 specification, structured tool results, resources, prompts, progress, stdio and Streamable HTTP. Earlier clients remain supported by the SDK's legacy protocol path.

Quick start

Python 3.11+ and uv are required.

uvx --from cite-nexus-mcp==0.2.3 cite-nexus-mcp --version
uvx --from cite-nexus-mcp==0.2.3 cite-nexus-mcp --no-env-file --list-providers

For an MCP client with a mcpServers configuration:

{
  "mcpServers": {
    "cite-nexus": {
      "command": "uvx",
      "args": ["--from", "cite-nexus-mcp==0.2.3", "cite-nexus-mcp", "--no-env-file"]
    }
  }
}

The client starts MCP over stdio. No API key or .env file is required. Running the server command directly waits for a client; it is not an interactive search shell. If your client cannot find uvx, use its absolute executable path. The first run downloads the package and its dependencies.

For a persistent installation, including the WTF-P companion:

uv tool install cite-nexus-mcp==0.2.3
cite-nexus-wtfp --check

The diagnostic starts the actual MCP service and lists providers without querying vendor APIs. See the WTF-P setup for the separate integration and executable-path setting.

For local HTTP clients:

uvx --from cite-nexus-mcp==0.2.3 cite-nexus-mcp --no-env-file --transport streamable-http --port 8000

Connect to http://127.0.0.1:8000/mcp. The CLI binds only to loopback and serves a single trust boundary. A public, multi-user deployment needs authentication, credential isolation and its own deployment integration; the CLI does not supply those features.

Sources and access

find-open-access can additionally enrich DOI records with Unpaywall when UNPAYWALL_EMAIL is configured. It returns reported legal OA locations and licensing information; it does not download PDFs or bypass access restrictions.

Free access is subject to provider rate limits and terms. Anonymous Semantic Scholar can be heavily throttled. OpenAlex access policies have changed over time; CiteNexus requires a key for that adapter even when an anonymous request happens to succeed. Institution keys do not by themselves guarantee subscription access or access outside an institution's network.

Commercial and metered adapters are called only when selected explicitly, when resolving their own namespaced IDs, or when deliberately configured as search defaults. Setting a key alone does not add a vendor to default searches. list-providers reports configuration, capabilities and documentation links without testing access or revealing key values.

Research workflows

Use CiteNexus to build a literature shortlist, repair an uncertain reference, check the source behind a citation count, or prepare a bibliography for your writing tool. For a full writing workflow, the WTF-P companion connects discovery to evidence tables, outlines and manuscript work while keeping candidate references visible for review.

Ask your assistant to use the tools directly, or start with the literature-review and bibliography-audit MCP prompts. Example tool arguments:

{"query": "reproducible machine learning benchmarks", "providers": ["crossref", "arxiv"], "limit": 5}

{"identifier": "10.1038/nature14539", "format": "bibtex"}

{"citation": "10.1038/nature14539", "expected_title": "Deep learning", "expected_year": 2015}

{"identifier": "PMID:26017442", "direction": "references", "provider": "europe_pmc", "limit": 10}

{"identifiers": ["10.1038/nature14539", "arxiv:1706.03762"], "format": "csl-json"}

search-papers defaults to 10 results per provider, not 10 overall. Results interleave each provider's ranking, then merge compatible records. limit is 1–50, or at most 20 when using SerpAPI or SearchApi. Use each provider's next_offsets with the same limit, selecting that provider for the next call. Europe PMC search uses next_cursors instead: pass its token as cursor, select only europe_pmc, keep the same query and limit, and leave offset at zero. Offsets must be multiples of the limit, up to 10,000. year_from, year_to and open_access_only filter the fetched page; unknown values are excluded. Abstracts are omitted from search results by default; set include_abstract: true or resolve a record to retrieve them. Search records are compact by default: field_sources is empty while sources still names every provider. Set detail: "full" or resolve a record for per-field provenance. Provider search languages differ: explicit advanced queries use that provider's syntax.

Supported identifiers include 10.1038/nature14539, arxiv:1706.03762, PMID:26017442, PMC12345, openalex:W2919115771, s2:<paperId>, s2:CorpusId:<number>, scholar:<numericClusterId>, scopus:<numericId>, and WOS:<recordId>. DOI, arXiv and PubMed URLs are recognized. Arbitrary URLs are not fetched, and unprefixed numeric strings are not assumed to be PubMed or Scholar IDs.

DOI resolution tries Crossref, DataCite and Europe PMC in order until a matching record is retrieved. An explicit providers list instead checks each compatible selected source and retains evidence and disagreements. Native identifiers route to their corresponding source. A title search always returns candidates; it never silently declares the first result a match.

Read the MCP resources cite-nexus://methodology, cite-nexus://providers and cite-nexus://workflows for machine-accessible guidance.

What “verified” means

Verification establishes that an identifier exists in retrieved metadata and checks any supplied title and year. It does not establish that a paper's findings are correct, that every author is present, that the paper supports a specific claim, or that no retraction exists. A provider outage produces unavailable, rather than a claim that the citation is fabricated.

Exports omit unavailable metadata. Partial author lists generate warnings; BibTeX uses others where appropriate. Citation counts retain their source, index and observation timestamp and are never summed across databases. Counts from different providers are not directly interchangeable. Metadata can contain errors or conflicting publication years, and an absent retraction flag is not proof that a paper is unretracted.

CSL-JSON is structured bibliographic data. Use a CSL processor or your reference manager to render APA, IEEE or a journal's house style. enhance-citation supports default and compact formatting; it does not infer keywords, awards, impact factors or missing authors.

Configuration

Copy .env.example to .env only if you need optional settings. Existing process environment variables take precedence. The CLI loads ./.env, or an explicit --env-file; it never searches parent directories. Importing the Python package does not read .env or call external services. Pass --no-env-file for clients that manage credentials themselves. The WTF-P companion and plugin manifests always use this mode.

Cached records retain their original retrieval and citation-observation timestamps. The cache is bounded and partitions responses by credentials. Responses larger than 8 MB are rejected. Long Retry-After cooldowns return a provider issue instead of holding the call open. Keys and upstream error bodies are excluded from tool errors. Local stderr logging keeps stdout clean for MCP.

See provider details, migration notes, and product direction for access nuances and boundaries.

Development and verification

To work from source:

git clone https://github.com/akougkas/cite-nexus-mcp.git
cd cite-nexus-mcp
uv sync --locked
uv run ruff check src tests scripts
uv run ruff format --check src tests scripts
uv run mypy src/cite_nexus_mcp
uv run python scripts/prepare_release.py --check
uv run python scripts/marketplaces.py --check
uv run pytest -q
uv build

The automated suite uses synthetic provider fixtures and mock HTTP, plus real local stdio and Streamable HTTP processes. It requires no vendor credentials and blocks external network access from the test process. Paid and institutional adapters have fixture-based contract coverage; live entitlement testing requires an account with access.

An explicit public-API smoke test is available separately:

uv run python scripts/smoke_public.py

It ignores account credentials and only contacts Crossref, DataCite, Europe PMC, arXiv and anonymous Semantic Scholar. Rate limits and network failures are reported as outcomes.

The official MCP Tasks extension is not advertised. SDK 2.2 does not implement it yet. batch-citations is a bounded synchronous workflow with progress and cancellation, not a durable job queue. The roadmap describes the criteria for adding persistent tasks, local libraries and further licensed providers.

The launch kit includes positioning, listing copy, a demo script, and promotional drafts. The marketplace inventory tracks 41 channels and leads, including verified routes, duplicate channels and inactive or unresolved sites. Source metadata, portable plugin manifests and image assets are prepared for review. See data handling before configuring optional providers or hosting.

License

MIT. Provider content, API access and redistribution remain subject to each source's terms.