Bible CLI

A Bible study operating system for AI agents — and the humans working with them.

bible is an offline command-line toolkit that lets any coding agent (Claude Code, Codex, opencode, or anything with a shell) do serious biblical study from the text itself: plain reading, original-language word study, morphology and clause-level syntax search ("who did what to whom"), cross-translation comparison, computed inner-biblical parallels, inductive reading sessions with an evidence-linked notebook, and canon-wide pattern analysis — with every claim traceable to a query instead of to training data.

$ bible interlinear "John 3:16" | head -6
John 3:16
Οὕτως     Houtōs    G3779  Adv       so
γὰρ       gar       G1063  Conj      For
ἠγάπησεν  ēgapēsen  G0025  V-AIA-3S  loved
ὁ         ho        G3588  Art-NMS   -
Θεὸς      Theos     G2316  N-NMS     God

$ bible grep-morph --stem niphal --tense participle --book Isaiah --count
98 matching words
  H0539 אָמַן  8
  H5375 נָשָׂא  7
  ...

$ bible freq --strongs H1285 --by-book     # where does "covenant" concentrate?
$ bible xref "Isa 53:5" --text             # how does the canon take this verse up?
$ bible compare "Rom 8:1" -t all           # where do translations diverge?

Everything is offline, deterministic, and open-licensed: two SQLite databases built from public-domain and CC-BY scholarly sources (see Data), downloaded once on first run.

Why

LLMs already "know" the Bible — approximately, unverifiably, and shaped by whatever their training data emphasized. Bible CLI gives an agent the opposite: a set of precise instruments over the actual text, so study becomes evidence-based. The agent reads the passage it cites, checks what a Greek or Hebrew word actually means by how it is actually used, counts patterns instead of gesturing at them, and reports what the text says as distinct from what it concludes. The bundled methodology (bible agent-setup) teaches exactly that discipline — a biblical theology workflow: plain reading → lexical analysis → context → canonical patterns → labeled synthesis.

Related MCP server: Bible MCP

Install

npm install -g @divine-creative-ministries/bible-cli

That's it — the scripture databases download automatically on first use (~90 MB, checksummed; set BIBLE_CLI_NO_AUTO_DOWNLOAD=1 to manage them manually with bible db download). Or try it with zero install:

npx @divine-creative-ministries/bible-cli passage "John 3:16"

For agents, add the study methodology to your project:

bible agent-setup claude          # this project: .claude/skills/bible-study/SKILL.md
bible agent-setup claude --user   # every session on this machine: ~/.claude/skills/
bible agent-setup generic         # appends a section to AGENTS.md (codex, opencode, …)

Or run it as an MCP server: bible mcp (stdio) exposes all commands as tools, and bible mcp --http serves the same tools over Streamable HTTP.

No install at all: a hosted endpoint runs at https://mcp.biblecli.org/mcp. Add it as a custom connector in the Claude apps (Settings → Connectors → Add custom connector) or in ChatGPT (Settings → Connectors → Advanced → Developer mode), and any chat can study Scripture with the full toolset — the study methodology is delivered to the model automatically via the MCP server's instructions.

Bring your own translation

Only openly-licensed translations can ship with the tool — but you can import a translation you have licensed access to (ESV, NIV, NASB, CSB, …) for personal study:

bible import ./esv-usfm --id ESV --name "English Standard Version"

Accepts a USFM folder, zip, or single file, or a TSV of <reference><TAB><text> lines. The import lands in a local-only database (bible-user.db) that is never uploaded, shared, or included in releases — the same posture as a music player playing your own files. Once imported, the translation works everywhere a translation id does: passage -t ESV, compare (included in all), read, full-text search, freq --word, and BIBLE_TRANSLATION=ESV makes it the default. Imported texts carry verse text only — the original-language layer (morphology, Strong's, interlinear) stays anchored to the shipped tagged data. Remove with bible import x --id ESV --remove.

Commands

Every command supports --json (machine-readable output), forgiving references (John 3:16-18, jn 3 16, 1jn2:5, Psalm 23, Gen 1:1-2:3), scoping (--book Isaiah, --book Gen-Deu, --book ot), and helpful errors (Unknown book 'Pslams'. Did you mean: Psalms…).

The data

All redistributable, all attributed (see bible licenses):

† The LXX ships as a separate optional artifact (bible db download-lxx) because its digitization carries CC BY-SA; the core and study databases stay public domain + CC BY only.

The databases are built by a reproducible pipeline in this repo (npm run pipeline) with a verification stage that checks canonical verse counts, versification edge cases (Psalm titles, Malachi 3/4, Joel 2/3), Ketiv/Qere handling, and morphology-code coverage — the build fails loudly rather than shipping silently wrong data.

Design notes

• Verse IDs are integers (BBCCCVVV; Gen 1:1 = 1001001) on a KJV-English spine; Hebrew/Greek versification differences are mapped at build time and preserved for display. Psalm superscriptions are verse 0.

• One row per morpheme: Hebrew prefixes/suffixes and Greek crasis components are individually tagged, so grep-morph and frequency counts are exact. Analytics default to the default text stream (Qere; NA-stream Greek) so textual variants never inflate counts.

• The CLI is the single source of truth; the MCP server shells into it, so both interfaces always agree.

• Quotation detection is computed, not curated, in three labeled confidence tiers — quotation (5+ identical words in a row), allusion (4+ word or lemma run), echo (gated 3-lemma runs or shared rare vocabulary; speculative by design and excluded from default output) — at two match levels: surface (verbatim) and lemma (inflection-independent, via a dictionary derived from the tagged NT; marked ≈). The lemma level is what catches Rev 1:7 ⇐ Dan 7:13. Purely thematic connections remain bible xref's domain.

• The methodology is a text-first protocol: three lanes matching evidence to the claim (lookup / word-passage / study), discovery-before-thesis via survey, mandatory falsification attempts, and provenance tags separating what is OBSERVED, what is PATTERN, and what is INFERENCE — with anything from outside the corpus labeled as interpretive tradition.

Known limitations (honest edges)

• --edition Greek texts are reconstructed from TAGNT's word-set + variant apparatus — accurate wording, but not a facsimile of a printed edition's punctuation or orthography.

• ~80 Aramaic Strong's numbers lack formal lexicon entries (gaps in the BDB mapping); their entries are synthesized from the tagged text's own glosses and labeled as such.

• English reverse lookup (bible word lovingkindness) is heuristic: lexicon glosses first, then which original words underlie verses containing the English word.

• The Swete LXX digitization lacks Ecclesiastes; Daniel uses Theodotion (which the NT normally follows).

Roadmap

• Semantic similarity via local embeddings

• Pericope/discourse boundaries for context-aware --context

• Versification traditions beyond Hebrew/Greek (Vulgate)

• Additional open-licensed translations, non-English included

• Syntax-aware search (MACULA clause/phrase trees)

PRs welcome.

License

Code: MIT. Data: see docs/DATA-SOURCES.md and bible licenses — public domain, CC BY 4.0, and CC0 in the core/study databases; the optional LXX artifact is CC BY-SA 4.0.