Bibverify

Bibverify

Verify bibliographic existence and metadata consistency without destructive edits.

Chinese · Install · Quick start · MCP · Development

[PyPI] [Python versions] [MIT License] [CI]

Bibverify is a BibTeX metadata verification tool for researchers, editors, automation, and AI assistants. It starts with exact DOI, PMID, PMCID, or arXiv identifiers and then scores candidates from title, author, year, venue, and pagination evidence.

Bibverify evaluates whether a bibliographic record can be located in the queried sources and whether its metadata agrees. It does not prove that research findings are true, data is authentic, or a venue is reputable. A missing database record is not evidence that a reference is fabricated. By default, the source .bib file is never overwritten.

Highlights

• Identifier-first lookup for DOI, PMID, PMCID, and arXiv IDs.
• Crossref, OpenAlex, Semantic Scholar, PubMed, Europe PMC, CORE, DBLP, arXiv, bioRxiv, and more.
• Explainable multi-signal matching across identifiers, title, authors, year, venue, and pages. A resolvable DOI with a materially different title becomes identifier_conflict instead of being hidden by title search.
• Structured provider outcomes distinguish a genuine no-match from rate limiting, authentication, network, parsing, and provider failures.
• Non-destructive merging preserves abstract, keywords, file, note, and custom fields that providers do not return; conflicting persistent identifiers are never overwritten automatically.
• A shared connection pool with retries, exponential backoff, and Retry-After support for 429/5xx responses.
• An expiring SQLite cache for successful GET responses; failures are never cached.
• Cross-platform path, encoding, Unicode filename, UTF-8 BOM, and CRLF handling.
• JSON output, stable exit codes, a Python API, and an MCP server built on the official SDK.
• Atomic output writes and byte-preserving backups.

Requirements

• Windows, macOS, or Linux
• Network access to the enabled metadata APIs
• Python 3.11–3.14 when using the Python distribution; npm, containers, and native packages bundle their runtime

GitHub Actions tests all three operating systems across all four supported Python versions.

Install

The current published release is v0.3.0. This branch prepares v0.4.0; commands marked v0.4.0 become usable only after their linked npm, GHCR, or GitHub Release artifact has been published.

Run without a permanent installation:

uvx bibverify --version

Starting with v0.4.0, Node.js users can use the zero-dependency npm launcher. It downloads the matching native release, verifies SHA256SUMS, and forwards every argument and exit code:

npx --yes @hylouis233/bibverify --versionpnpm dlx @hylouis233/bibverify --versionbunx --bun @hylouis233/bibverify --version

For a persistent CLI in an isolated Python environment, use uv tool or pipx:

uv tool install bibverify

pipx install bibverify

Inside a virtual environment, regular pip also works:

python -m pip install --upgrade bibverify

Each release also provides smoke-tested native packages on GitHub Releases. Beginning with v0.4.0, the release matrix covers Windows x64 plus macOS and glibc 2.28+ Linux on both x64 and ARM64. musl-based Linux users should use the Python package or container image. Windows ARM64 is not published natively yet because an MCP runtime dependency does not currently provide Windows ARM64 wheels; npm automatically uses the tested x64 build under Windows 11 emulation, or you can use the ARM64 container instead.

The same release publishes a multi-architecture container:

docker run --rm ghcr.io/hylouis233/bibverify:0.4.0 --version

Package-manager manifests are generated from the final release bytes, not from unverified build inputs. Homebrew, Scoop, and WinGet are not live catalog entries yet: v0.4.0 will attach submission- ready manifests to GitHub Release, after which each external catalog still requires onboarding or review. See Distribution channels for rollout status.

Quick start

Convert a DOI to BibTeX

bibverify doi 10.1038/nature12373 --key example2013

For a machine-readable response:

bibverify doi 10.1038/nature12373 --json

Verify a .bib file

Create a starter configuration:

bibverify config init

Place references.bib next to the configuration and run:

bibverify check --config config.json

You can override the input and output paths from the command line:

bibverify check references.bib --config config.json --output-dir bibverify-output

Inspect results without writing any file:

bibverify check references.bib --dry-run --json

After reviewing the report, explicitly apply high-confidence field updates. Bibverify creates a byte-for-byte backup first:

bibverify check references.bib --apply

PowerShell example:

py -m bibverify check '.\Bibliography\references.bib' --output-dir '.\Verification results'

The v0.2 forms remain available for compatibility, although new scripts should use subcommands:

bibverify config.jsonbibverify --doi 10.1038/nature12373 --key example2013

Configuration

A minimal configuration looks like this:

{  "language": "EN",  "bib_file": "references.bib",  "encoding": "auto",  "output_dir": "bibverify-output",  "user_info": {    "email": "[email protected]",    "app_name": "Bibverify"  }}

See config_template.json for every commonly used option.

Path behavior is intentionally predictable:

• Relative bib_file and output_dir values are resolved from the directory containing config.json, not the shell's current directory.
• If output_dir is omitted, output is written next to the input bibliography.
• encoding: "auto" tries UTF-8 with BOM, UTF-8, and GB18030. It does not fall back to Latin-1 and silently turn unknown bytes into mojibake.

API keys and email

Keys can be stored in a local configuration, but environment variables are safer and harder to commit accidentally:

PowerShell:

$env:BIBVERIFY_EMAIL = '[email protected]'$env:BIBVERIFY_OPENALEX_API_KEY = '...'bibverify check --config config.json

Bash or Zsh:

export BIBVERIFY_EMAIL='[email protected]'export BIBVERIFY_OPENALEX_API_KEY='...'bibverify check --config config.json

Query and matching settings

{  "query_settings": {    "delay_between_requests": 0.5,    "timeout": 10,    "connect_timeout": 3.05,    "read_timeout": 20,    "max_retries": 3,    "backoff_factor": 0.5,    "stop_on_first_match": true,    "match_threshold": 0.86,    "ambiguous_threshold": 0.68,    "auto_update_threshold": 0.92,    "cache_enabled": true,    "cache_ttl_hours": 168,    "cache_path": ".bibverify-cache.sqlite3"  }}

connect_timeout and read_timeout separately bound connection setup and response reads; the compatibility timeout field remains available. match_threshold controls automatic candidate acceptance, ambiguous_threshold controls which plausible candidates enter review, and auto_update_threshold is an additional gate for field changes. Higher values are more conservative. Relative cache paths are resolved from the config directory.

The official bioRxiv details route does not provide arbitrary title search. Bibverify therefore calls bioRxiv only for an exact 10.1101/... DOI and leaves title-only discovery to providers whose contracts support it, such as Crossref and Europe PMC.

CLI reference

bibverify check [BIB_FILE] [--config PATH] [--output-dir DIR] [--format txt|json|jsonl|csv] [--dry-run|--apply] [--json]bibverify doi DOI [--key KEY] [--config PATH] [--json]bibverify config init [--output PATH] [--force]bibverify doctor [--config PATH] [--json]bibverify providers list [--json]bibverify cache clear [--config PATH]bibverify benchmark [--dataset PATH]bibverify mcp [--config PATH] [--workspace-root DIR] [--transport stdio|streamable-http]bibverify agent init [--target generic|codex|claude|cursor] [--output PATH]bibverify skill export [--target ...]

Exit codes:

With --json, stdout contains JSON only. Diagnostics go to stderr, which keeps the command safe for CI and scripts.

Output files

For an input named references.bib, Bibverify may create:

• bibverify_report_<timestamp>.<format>: complete states, candidates, provider errors, confidence, and field provenance in txt, json, jsonl, or csv.
• references_backup_<timestamp>.bib: byte-for-byte copy of the source.
• references_updated_<timestamp>.bib: complete bibliography after non-destructive merging; omitted when nothing changes.
• references_review_<timestamp>.bib: ambiguous, not-found, unavailable-source, identifier-conflict, or invalid entries; omitted when nothing needs review.

The report-level complete value is true only when every entry completed verification. A rate-limited or unreachable provider makes it false even if another source found a candidate. Each field_diffs record includes original/suggested values, source, confidence, normalized equivalence, action, and reason.

Each output can be disabled independently through output_settings. --dry-run overrides those settings and performs zero writes. The default command only writes proposals; only --apply changes the source after a backup.

Provider ranking

Static priority is only the starting point:

1. A DOI promotes Crossref and uses its exact endpoint first. A resolvable DOI with a materially different title stops as identifier_conflict.
2. PMID, PMCID, or biomedical hints promote PubMed and Europe PMC.
3. An arXiv identifier promotes arXiv.
4. Computer-science venue hints promote DBLP.

Unpaywall is currently treated as open-access enrichment, not as a primary bibliographic metadata provider. Provider states distinguish matched, no_match, ambiguous, rate_limited, auth_error, network_error, parse_error, provider_error, and skipped.

MCP and AI assistants

Bibverify uses the official MCP Python SDK and supports local stdio and Streamable HTTP transports.

Start a stdio server:

bibverify mcp --config config.json --workspace-root .

MCP client configuration:

{  "mcpServers": {    "bibverify": {      "command": "bibverify",      "args": ["mcp", "--config", "config.json"]    }  }}

The npm launcher can be used directly by Node-oriented MCP clients:

{  "mcpServers": {    "bibverify": {      "command": "npx",      "args": ["--yes", "@hylouis233/bibverify", "mcp"]    }  }}

Containerized stdio MCP keeps the current directory as the only writable workspace:

docker run --rm -i -v "$PWD:/workspace" ghcr.io/hylouis233/bibverify:0.4.0 \  mcp --workspace-root /workspace

Start Streamable HTTP:

bibverify mcp --transport streamable-http --config config.json

MCP treats the configuration directory as its workspace root by default. It rejects config or .bib reads outside that root and blocks report, cache, or update writes outside it. Broaden access only by explicitly setting --workspace-root when starting the server. The official MCP SDK handles protocol negotiation, schemas, structured results, progress, and cancellation.

Available tools:

• doi_to_bibtex
• rank_lookup_sources
• explain_update_diff
• verify_bib_file

Generate setup files for Codex, Claude, Cursor, or a generic MCP client:

bibverify agent init --target codex --output .bibverify-agent --config config.jsonbibverify doctor --config config.json

Distribution channels

The npm package is a small launcher rather than a second implementation. The verification engine remains the same Python codebase across PyPI, npm, native packages, and containers. Release assets include SHA256SUMS; npm verifies the selected binary before execution, and Homebrew, Scoop, and WinGet manifests pin the same release hashes.

Python API

from bibverify.checker import BibTeXChecker
checker = BibTeXChecker("config.json")summary = checker.run()print(summary["counts"])

from bib_check import BibTeXChecker remains compatible during the 0.3 release line, but new code should use the package import above.

Development

git clone https://github.com/Hylouis233/bibverify.gitcd bibverifypython -m venv .venv

After activating the environment:

python -m pip install -e ".[dev]"python -m pytestpython -m ruff check src tests tools bib_check.pypython -m ruff format --check src tests tools bib_check.pypython -m mypypython -m buildpython -m twine check dist/*python -m bibverify benchmark --dataset benchmarks/cases.jsonpython -m pip_audit . --strict(cd npm && npm ci && npm test && npm pack --dry-run)docker build --tag bibverify:dev .

Before pushing, run python tools/fix.py. It applies ruff format and import sorting to exactly the paths CI checks, then runs CI's ruff check and ruff format --check commands and fails on anything that needs a manual fix. pre-commit install runs the same command as a commit hook (see .pre-commit-config.yaml).

CI tests Windows, macOS, and Linux on Python 3.11–3.14, including provider fixtures, golden write-safety tests, lint, typing, coverage, the offline benchmark, dependency auditing, Python and npm package builds, a container smoke test, and a CycloneDX SBOM. GitHub Actions are pinned to commit SHAs, while the MCP Publisher is version-pinned and SHA-256 verified. PyPI and npm use Trusted Publishing with provenance attestations, and GHCR publishes multi-architecture images with an SBOM and build provenance.

benchmarks/cases.json is a small offline regression set covering short-title false positives, DOI conflicts, preprint title variants, Unicode/LaTeX, and fabricated author combinations. It is not a complete scientific evaluation and its scores must not be interpreted as real-world performance ceilings. Contributions of broader, redistributable, human-labeled cases are welcome.

Citation

If Bibverify supports your research, please cite:

@software{bibverify2025,  title = {Bibverify: A Multi-Platform BibTeX Reference Verification Tool},  author = {Hong Liu},  year = {2025},  url = {https://github.com/Hylouis233/bibverify},  doi = {10.5281/zenodo.17338090}}

License

MIT License