bible-mcp

An MCP server for Christian scholarship and research. Non-commercial, aiming to become a public resource. See corpus-survey.md for the full source/license landscape and ROADMAP.md for direction.

What's in it (v0.9)

All in one SQLite file (db/bible.db, ~230MB) with FTS5 full-text search. Prose works are addressed as WORK.chapter.paragraph. See DESIGN.md for the architecture rationale.

Related MCP server: mcp-bible

Tools

get_passage · search (stemmed BM25, phrases, AND/OR/NOT; version takes BSB, WEB, or any prose work id) · semantic_search (meaning-based, hybrid RRF with keyword search) · find_similar (nearest passages to any verse or paragraph, across layers) · word_study (by Strong's — zero-padding optional, lettered homograph variants surfaced together — or by lemma, pointed or unpointed, or English gloss) · get_interlinear (versemap-corrected across all books, not just Psalms) · get_cross_references (with target text) · get_citations (patristic citations of a verse; footnote-extracted, so sparse — full-text search is the thorough probe) · get_entity · entities_in_passage · read_work · compare_versions · corpus_info

Prompts

The server also exposes the synthesis pipeline as MCP prompts, generated from .claude/skills/ (the source of truth): corpus_survey(theme) — Layer 0 raw reading discipline; corpus_composer(theme) — Layer 2 research-brief composition. Any MCP client that installs bible-mcp gets the discipline bundled with the data.

Setup

pip install -r requirements.txt   # core: just `mcp`
pip install fastembed numpy       # optional semantic tier (semantic_search, find_similar)
python3 scripts/build_db.py       # rebuild db from data/sources (optional; db ships built)
python3 scripts/embed.py          # regenerate embeddings (optional; ships built; resumable)

The first semantic query downloads the embedding model (~65MB, one time). Without fastembed/numpy installed, all non-semantic tools work normally. See DESIGN.md for architecture decisions (model choice, chunking, hybrid retrieval).

Claude Desktop / Cowork — add to MCP config:

{
  "mcpServers": {
    "bible-mcp": {
      "command": "python3",
      "args": ["/path/to/bible-mcp/server.py"]
    }
  }
}

Layout

data/sources/   raw acquired sources (immutable; re-derive, never edit)
data/versemap.tsv        derived MT/NA <-> English verse alignments
data/additions-*.json    prepared incremental ingestions (applied via scripts/ingest_additions.py)
scripts/        schema.sql, lib_refs.py, build_db.py, build_versemap.py, align_splits.py, ingest_additions.py
db/bible.db     the built corpus (SQLite + FTS5)
outputs/        synthesis-pipeline artifacts (Layer 0 surveys, Layer 2 briefs, reports)
.claude/skills/ corpus-survey + corpus-composer (pipeline skills; source for MCP prompts)
server.py       MCP server (FastMCP, read-only on the db)

Public edition

This is a non-commercial public resource. Licensing is layered — sources keep their own licenses (all PD / CC BY / CC BY-SA); the code is PolyForm Noncommercial 1.0.0; the compilation and derived data (versemap, citation graph, embeddings, outputs layer) are CC BY-NC 4.0. Full details: LICENSE.md, attributions in NOTICE.md, rationale in DECISIONS.md.

Three ways to use it:

1. Remote (no install) — add the hosted endpoint as a custom connector in any MCP client (Claude: Settings → Connectors → Add custom connector). URL in the registry listing / repo description.

2. Local (stdio) — clone, download db/bible.db from the latest GitHub Release (or rebuild from data/sources/), add the config above.

3. Self-host — Dockerfile ships the Streamable-HTTP server; see DEPLOY.md.

Design notes

• Canonical addressing: OSIS-style refs (Gen.1.1) everywhere; every table keys on them.

• Provenance: every document row carries source, license, and license tier (A=free, B=share-alike, C=non-commercial, D=closed — we are non-commercial, so C is usable).

• The 16 verses "missing" from BSB (Matt 17:21, John 5:4, Acts 8:37…) are the standard critical-text omissions, not bugs.

• Versification: versemap (all books) + psalm_offsets (superscriptions) align MACULA's MT/NA numbering to English. Both derived empirically from the corpus's own two witnesses and validated by gloss-text overlap; the table schema is TVTMS-compatible so the STEPBible data can extend it to other traditions (LXX, Vulgate) later.

• Ignatius ships in the shorter (Middle) recension only; the ANF longer recension is a later expansion and would duplicate every chapter in search.

• The research outputs layer is generated, clearly tagged (layer='output'), and can never be confused with primary sources; draws_on links tie each artifact to its grounding refs.

• Next per ROADMAP.md: Irenaeus (source acquisition), tier-2 citation shingling (three confirmed footnote-index misses as fixtures), TVTMS proper.