CiteNexus MCP
CiteNexus MCP is a scholarly discovery and citation service that searches open and licensed academic sources and produces traceable, source-backed citation exports.
Search papers concurrently across Crossref, DataCite, Europe PMC, arXiv, Semantic Scholar, OpenAlex, Google Scholar vendors, Scopus, and Web of Science, with filters and pagination.
Resolve DOIs and namespaced identifiers (arXiv, PMID, PMC, OpenAlex, Semantic Scholar, Scholar, Scopus, WOS).
Export deterministic BibTeX, RIS, or CSL-JSON citations with provenance and warnings.
Batch-resolve and export 1–20 identifiers with ordered results, partial failures, and progress.
Verify citations against sources, comparing supplied title/year and distinguishing mismatch, ambiguity, absence, and provider outage.
Normalize BibTeX formatting without inventing bibliographic facts.
Traverse citations and references from provider graphs.
Report observed citation counts per provider, without summing across databases.
Find source-reported open-access locations and licenses, optionally enriched by Unpaywall.
Discover legacy Google Scholar cluster candidates via SerpAPI or SearchApi.
List providers, access tiers, capabilities, configured availability, and required environment variables without exposing secrets.
Allows searching and retrieving academic citation metadata from Google Scholar, including finding cluster IDs, generating BibTeX entries, and fetching paper metrics.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@CiteNexus MCP Get BibTeX for the paper 'Attention is all you need'"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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 0.2.5 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==0.2.5 cite-nexus-mcp --version
uvx --from cite-nexus-mcp==0.2.5 cite-nexus-mcp --no-env-file --list-providersFor an MCP client with a mcpServers configuration:
{
"mcpServers": {
"cite-nexus": {
"command": "uvx",
"args": ["--from", "cite-nexus-mcp==0.2.5", "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==0.2.5
cite-nexus-wtfp --checkThe 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==0.2.5 cite-nexus-mcp --no-env-file --transport streamable-http --port 8000Connect 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.
Clio
Declare CiteNexus in the user scope ~/.config/clio-coder/mcp.yaml. Every CiteNexus tool is
read-only, and searches contact the selected providers, so the server needs the network grant:
version: 1
servers:
- id: cite-nexus
command: /absolute/path/to/cite-nexus-mcp # from `uv tool install cite-nexus-mcp==0.2.5`
args:
- --no-env-file
timeoutMs: 120000
actionClass: read
network: true
toolActionClasses:
batch-citations: read
enhance-citation: read
find-open-access: read
find-scholar-id: read
get-citation: read
list-providers: read
paper-metrics: read
related-papers: read
resolve-paper: read
search-papers: read
verify-citation: readTool names are hyphenated, exactly as tools/list reports them. Run
clio-coder mcp test cite-nexus --call list-providers to check the declaration. Clio honors
network: true only in a user declaration. A plugin-scope declaration's own network field is
ignored, so grant it explicitly with clio-coder mcp trust cite-nexus --network.
Small local models should call search-papers with detail: "minimal" and limit_total.
Related MCP server: google-scholar-search-mcp
Sources and access
Provider ID | Coverage | Access | Setup |
| Publisher-deposited DOI metadata and citation counts | Free; default | Optional |
| Datasets, software and other DOI-registered research outputs | Free; default | None |
| Biomedical literature, PubMed/PMC records, references and citations | Free; default | None |
| Preprints in physics, mathematics, computing and related fields | Free | Select explicitly; paced to one request per three seconds |
| Academic Graph discovery and citation traversal | Free/shared anonymous limits; optional key | Optional |
| Multidisciplinary research graph | Free credits / metered account access |
|
| Google Scholar through SerpAPI vendor API | Commercial; account allowances vary |
|
| Google Scholar through SearchApi.io vendor API | Commercial; account allowances vary |
|
| Elsevier's curated literature index | Institutional/commercial entitlement |
|
| Clarivate Web of Science Starter | Institutional/commercial entitlement |
|
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"}Tool | Purpose |
| Inspect source capabilities, access tiers and credential configuration |
| Concurrent federated search, conservative deduplication and per-source pagination |
| Retrieve and validate a DOI or namespaced identifier; optionally merge selected sources |
| Export one resolved record as |
| Check that an identifier or BibTeX entry exists and matches supplied title and year; never checks claim support |
| Resolve and export 1–20 identifiers with ordered results, partial failures and progress |
| Page through a source's citations or references |
| Return observed citation counts, separated by provider and index |
| Find source-reported OA locations, with optional Unpaywall enrichment |
| Normalize BibTeX formatting while preserving supplied bibliographic facts |
| Legacy Scholar discovery through SerpAPI or SearchApi; returns candidates for review |
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. Ordering interleaves each
provider's own ranking, so the first record is the first selected provider's top hit. For
computer-science topics, select arxiv and semantic_scholar; plain arXiv queries require every
non-stopword term, while explicit ti:, abs:, AND or OR syntax passes through unchanged. 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. For small context
windows, detail: "minimal" returns only id, title, year, the first three authors, venue and
DOI/arXiv/PMID, and limit_total caps records across all providers:
{"query": "burst buffers", "limit": 5, "limit_total": 5, "detail": "minimal"}.
A record without an abstract carries abstract: null and abstract_missing_reason.
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.
verify-citation reports match: "identifier" when only the identifier was checked and
match: "metadata" when a supplied title or year also matched; claim_support is always
not_checked.
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.
Variable | Default | Meaning |
|
| Explicit search source IDs, comma separated |
|
| Timeout in seconds for each HTTP request |
|
| Time budget for a provider operation or identifier resolution |
|
| Retries for transport failures, HTTP 429 and 5xx responses |
|
| Maximum simultaneous outbound requests |
|
| In-process response-cache lifetime in seconds; |
|
| Maximum cached responses; |
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 buildThe 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.pyIt 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.
Available Tools
11 toolsbatch-citationsARead-onlyIdempotent
Resolve and export up to 20 identifiers, preserving order and per-item failures. Reports progress; no durable background task is created. Example: {"identifiers": ["10.1038/nature14539", "arXiv:1706.03762"], "format": "bibtex"}
Inputs: identifiers (required), format, providers. Returns: items, failed, completed
enhance-citationARead-onlyIdempotent
Normalize BibTeX formatting while preserving supplied metadata. Does not invent authors, awards or keywords. Use verify-citation to check facts. Example: {"bibtex": "@article{key, title={A work}, year={2020}}"}
Inputs: bibtex (required), template. Returns: bibtex, warnings
find-open-accessARead-onlyIdempotent
Find source-reported open-access locations, with optional Unpaywall enrichment. Returns links and licenses without downloading full texts. Example: {"identifier": "10.1371/journal.pone.0000308"}
Inputs: identifier (required), providers. Returns: paper, issues, locations
find-scholar-idARead-onlyIdempotent
Legacy Google Scholar discovery via paid SerpAPI or SearchApi. Returns candidates with namespaced cluster IDs; the first result is not identity proof. Requires SERPAPI_API_KEY or SEARCHAPI_API_KEY. Example: {"query": "attention is all you need"}
Inputs: query (required). Returns: query, issues, papers, returned, +4 more
get-citationARead-onlyIdempotent
Resolve a DOI or namespaced ID and export deterministic BibTeX, RIS or CSL-JSON with provenance and missing-field warnings. Legacy scholar_id remains accepted. Example: {"identifier": "10.1038/nature14539", "format": "bibtex"}
Inputs: format, providers, identifier, scholar_id. Returns: paper, format, citation, warnings
list-providersARead-onlyIdempotent
List sources, access tiers, configured availability, supported operations and setup variables. Does not call vendors or expose secrets. Example: {}
Inputs: no named parameters. Returns: policy, providers
paper-metricsBRead-onlyIdempotent
Retrieve observed citation counts by provider and index, never an estimated impact factor or sample length. Select multiple providers to compare coverage. Example: {"identifier": "10.1038/nature14539", "providers": ["crossref", "openalex"]}
Inputs: providers, identifier, scholar_id. Returns: paper, issues, metrics, interpretation
resolve-paperARead-onlyIdempotent
Resolve a DOI, arXiv ID, PMID, PMCID, OpenAlex, Semantic Scholar, Scholar, Scopus or WOS ID. Validates returned identifiers; use search-papers for titles. A missing abstract is returned as abstract: null with abstract_missing_reason; never infer findings from the title. Example: {"identifier": "10.1038/nature14539"}
Inputs: identifier (required), providers. Returns: paper, issues, identifier
search-papersARead-onlyIdempotent
Search scholarly sources concurrently (defaults: Crossref, DataCite, Europe PMC; for computer science select ["arxiv", "semantic_scholar"]). Ordering interleaves each provider's own ranking, so the first record is the first provider's top hit, not a global best match. limit is PER PROVIDER: limit=5 with 3 providers returns up to 15 records. limit_total caps records returned across all providers. detail=minimal returns only id, title, year, first 3 authors, venue and doi/arxiv/pmid (about 5x smaller than compact); compact drops per-field field_sources; full keeps everything. Page one source at a time using next_offsets, or next_cursors as cursor for Europe PMC, with the same query and limit. Filters apply to fetched pages. Partial failures remain visible. Results are candidates; resolve-paper an identifier before citing. Example: {"query": "burst buffers HPC I/O", "limit": 5, "limit_total": 5, "detail": "minimal"}
Inputs: query (required), limit, cursor, detail, +7 more. Returns: query, issues, papers, returned, +4 more
verify-citationARead-onlyIdempotent
Check that a citation's identifier exists and, if expected_title/expected_year are given, that they match the source record. Pass a DOI/arXiv/PMID or one BibTeX entry as citation (identifier is accepted as an alias). status=verified with match=identifier means the ID resolved; match=metadata means the supplied title/year also matched. It NEVER checks whether the paper supports a claim (claim_support is always not_checked). Title-only input returns ambiguous candidates. Provider outages are distinct from missing records. Example: {"citation": "10.1038/nature14539", "expected_title": "Deep learning", "expected_year": 2015}
Inputs: citation, providers, identifier, expected_year, +1 more. Returns: match, paper, issues, status, +5 more
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
There is a dense cluster of resolution/export tools — resolve-paper, get-citation, verify-citation, and batch-citations — that all take an identifier and return a source record, so an agent must read carefully to pick the right one (get-citation for formatted export, verify-citation for title/year matching, batch for up to 20). The descriptions do draw these distinctions (match semantics, format output, bulk limits), but the boundaries are subtle rather than obvious. The remaining tools (search-papers, find-open-access, paper-metrics, related-papers, list-providers, find-scholar-id) are clearly distinct.
All names use consistent kebab-case and are predominantly verb-led (resolve-paper, search-papers, verify-citation, enhance-citation, find-open-access, get-citation, list-providers). A few deviate to noun-led forms (batch-citations, paper-metrics, related-papers), which is a minor inconsistency but still readable and predictable.
Eleven tools is well within a healthy range for a scholarly-citation server and each covers a plausible workflow step. However, the overlap among the four resolution/export tools suggests a couple could be consolidated, making it slightly heavier than strictly necessary.
The surface covers discovery, resolution, verification, bulk handling, formatting/enhancement, open-access lookup, metrics, related-work paging, and provider introspection — a solid lifecycle for the citation domain. Minor gaps remain: batch-citations accepts only identifier lists (no BibTeX/file input), and there is no cross-provider deduplication or full-text retrieval.
Maintenance
Related MCP Connectors
Google Scholar papers, citations and author profiles. Paid, on Apify's hosted MCP.
Citations, abstracts, PDFs, as one MCP tool through Apify
Web search, scraping, RAG answers with citations, and translation as MCP tools.
Federated search of books and papers, BibTeX/RIS citations, open-access retrieval and reading.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenance🔍 Enable AI assistants to search and access Google Scholar papers through a simple MCP interface.435-
- AlicenseAqualityBmaintenanceAn MCP server for searching Google Scholar, enabling paper search, author lookup, citation tracking, and BibTeX export for AI assistants and automation workflows.52MIT
- AlicenseAqualityDmaintenanceEnables AI assistants to search and analyze Google Scholar publications, authors, citations, and download papers.112MIT
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to search and access academic papers from Google Scholar via MCP protocol, with support for keyword, author, and year filtering, CAPTCHA handling, and BibTeX retrieval.3MIT