Independent Project Not affiliated with, sponsored by, or endorsed by the Watch Tower Bible and Tract Society or Jehovah's Witnesses.
jw-agent-toolkit
ES

Agentic Toolkit · Open source · Python 3.13+

A technical library for the public content of jw.org

A Python monorepo with HTTP clients, deterministic parsers, MCP server, hybrid RAG, procedural agents, and local fine-tuning. Built so any developer can talk with biblical content without duct-taping HTML.

Start in 4 lines View architecture ⌘ git clone
python — jw-core 
>>> from jw_core import parse_reference
>>> ref = parse_reference("Juan 3:16")
>>> ref.display()
'John 3:16'
>>> ref.wol_url(lang="es")
'https://wol.jw.org/es/wol/b/r4/lp-s/nwt/43/3#study=discover&v=43:3:16'
>>> ref.book.code, ref.chapter, ref.verses
('JHN', 3, [16])

# Determinístico. Sin LLM. Sin red (cassetes pytest).
# Cita verificable en wol.jw.org · 7 idiomas tier-1.
Python Packages
10
Tests passing
2.99k +
MCP Tools
135 +
ASR languages
1672

Recommended

Three ways to start

Guide · Start

Resolve any biblical citation in 4 lines of Python

From the string 'John 3:16' to a canonical wol.jw.org URL in any supported language. No LLM, no network in tests, 100% verifiable.

Concept · Architecture

Why there is no LLM on the critical path

Deterministic parsers, always verifiable citations, decoupled layers. AI synthesis is optional and lives in Claude Desktop, not in the toolkit.

Guide · Advanced

Train your own local model with your publications

Complete 5-phase pipeline: JWPUB extraction → dedup → synthetic Q&A → LoRA → GGUF export. Unsloth stack on local GPU, sending nothing to the cloud.

Monorepo · uv workspace

Ten packages, a single dependency: jw-core.

Dependencies flow downwards: every package depends on jw-core, nothing else. This allows using the CLI, MCP, RAG or agents independently or composed.

Paquete · 01

jw-core

El núcleo determinístico

Librería principal: 6 clientes HTTP, 10 parsers, writers JWPUB/.jwlibrary, 6 providers ASR (Omnilingual 1672 idiomas + whisperX diarizado), NLLB-200, schemas organized-app, voiceprints, infra F9 + multimodal end-to-end (talk-lab F68, broadcasting/visual F69, image-quote F70, book-camera F71), drift diacrónico F72 y voz familiar consentida F76.

  • 6 clientes · 10 parsers · 3 writers · 17 locales
  • ASR 1672 idiomas + whisperX diarizado
  • Multimodal: talk-lab · visual index · image-quote · book-camera
Paquete · 02

jw-cli

Terminal para mortales

CLI Typer + Rich con 13+ comandos top-level (verse · search · daily · download · jwpub build · library · omnilingual · translate · transcribe · …). Wrapper directo sobre jw-core con output Rich o JSON.

  • 13+ comandos: verse · search · daily · …
  • F55: jw jwpub build · library · omnilingual · translate
  • Typer + Rich · 7 idiomas
Paquete · 03

jw-mcp

Puente con tu agente

Servidor Model Context Protocol que expone ~135 herramientas a Claude Desktop, Claude Code o cualquier cliente MCP. Cubre F57-F76: meeting, brain, ingest, ASR diarizado, memoria, meta-orchestrator, sparring, reasoner, talk-lab, broadcasting visual, book-camera (con REST endpoints), drift, voice-clone.

  • ~135 tools sobre stdio MCP
  • FastMCP · Claude/IDE compatible
  • RAG + agentes + brain + meeting + agéntica F65-F76
Paquete · 04

jw-rag

Recuperación híbrida

Indexación vectorial + búsqueda híbrida BM25 + cosenos + Reciprocal Rank Fusion. Sobre Biblia, publicaciones, EPUB y JWPUB descifrados.

  • BM25 + vector + RRF
  • Sync incremental por source_id
  • Embeddings configurables
Paquete · 05

jw-agents

Orquestación multipaso + agéntica verificable

Agentes procedurales determinísticos + meta-orchestrator F65 con planner+critic NLI, conversation-sparring F66 con 6 personas y doctrinal-reasoner F67 con ReAct + golden set. Sin LLM en el camino crítico.

  • 4 agentes procedurales + meta-orchestrator
  • Conversation sparring · 6 personas
  • Reasoner ReAct con NLI crítico
Paquete · 06

jw-finetune

Tu modelo, tus datos

Plataforma local estilo Unsloth Studio: extrae JWPUB/EPUB, genera Q&A sintéticos, entrena LoRA y exporta a GGUF/MLX. Sin distribuir pesos.

  • Extract → Synth → Train → Export
  • Unsloth · LoRA · GGUF · MLX
  • Textual TUI + WebSocket monitor
Paquete · 07

jw-eval

Red de seguridad doctrinal

Suite de evaluación con regresión: 3 capas (estructural, citas, semántico) que gating cada PR contra 47 golden Q&A. Convierte 'confío en mí' en métrica.

  • 47 golden cases · 6 agentes
  • L1 estructural · L2 citas · L3 semántico
  • Embeddings + LLM judge híbrido
Paquete · 08

jw-gen

Generación con difusión, uso personal

Octavo paquete del monorepo. Genera imagen / audio / video para presentaciones y discursos personales. Watermark obligatorio + metadata EXIF/XMP + safety filters anti-emulación-JW-oficial.

  • Image · Audio · Video — APIs SOTA
  • Watermark + EXIF + disclaimer obligatorio
  • 3 safety filters no-negociables
Paquete · 09

jw-brain

Second-brain + Bible Knowledge Graph

Karpathy-style second-brain con compiler dual-backend (DuckDB/Neo4j) + Wiki sobre Obsidian. F58 añadió BibleKnowledgeGraph JW-puro: 250 personas, 150 lugares con geocoordenadas, 10 periodos según cronología JW (607 a.E.C. para Jerusalén), CLI jw brain {init, compile, query, lint, import-bible, learn-headwords}.

  • GraphRAG DuckDB + Neo4j
  • BibleKG JW-puro (607 a.E.C.)
  • 475 personas + 259 lugares + 16 geocoords
Paquete · 10

jw-meeting-media

Reunión-en-vivo · clean-room

Descubrimiento programa semanal mwb/w desde WOL, descarga media (imágenes/videos/audio/JWPUB), presenter Tauri con drag-drop + monitor externo automático + multi-congregación. Implementación clean-room (NO portada del repo M³ AGPL-3.0).

  • Clean-room (no port AGPL)
  • Presenter Tauri + drag-drop
  • Monitor externo automático
Paquete · 11

jw-interp

Interpretabilidad mecanicista tri-modelo

Probing lineal por principio, steering vectors + activation patching, Qwen-Scope adapter (TopK residual 32k features) y Gemma Scope wrapper (JumpReLU SOTA residual + MLP + attention). Runtime probe store + fidelity_wrap Tier 4 observacional sin acoplamiento jw-agents → jw-interp. Paquete entregado en F80.

  • Probing lineal por principio (sklearn)
  • Steering vectors + activation patching pure-numpy
  • Qwen-Scope (TopK) + Gemma Scope (JumpReLU) cross-family
Paquete · 12

jw-meeting-scheduler

Solver de asignaciones + importer organized-app

Paquete F81: importer organized-app JSON backup con CRDT diff + dry-run, store SQLite cifrado por congregación con FieldEncryptor + salt PBKDF2, CLI de edición manual del roster (people list / person edit / history), constraints YAML Pydantic strict por congregación, y CP-SAT solver (F81.3, pendiente) que produce ProposedSchedWeek con infactibilidad estructurada.

  • organized-app importer · dry-run + CRDT diff
  • FieldEncryptor + salt PBKDF2 por congregación
  • CLI: people list · person edit · history · constraints
Paquete · 13

jw-legal

BrainDomain plugin: casos legales TJ vs Estado

Plugin F82.1 que registra el BrainDomain legal-cases-tj en jw-brain via entry-point jw_agent_toolkit.brain_domains. 6 NodeTypeSpec (LegalCase, Law, Territory referenciando jw_core.territories sin duplicar, CourtPrecedent, LegalArgument, PersecutionEvent) + 8 EdgeTypeSpec con CONTRADICTS non-directional sensitive (la política de conflictos del segundo cerebro flag en vez de fusionar).

  • Plugin via entry-point jw_agent_toolkit.brain_domains
  • 6 NodeTypeSpec + 8 EdgeTypeSpec
  • Territory referencia jw_core.territories sin duplicar

Architecture

Strict layers.
Verifiable citations.

  • jw-core imports nothing from the workspace — it is the clean foundation.
  • jw-mcp is the only coupling point that binds the global RAG.
  • Every response from any agent includes a canonical URL to wol.jw.org.
  • Parsers are deterministic. AI synthesis happens outside the toolkit.
Read the architecture manual 
┌──────────────────────────────────────────────────────────────┐
│  Skills (Markdown)      Agentes (orquestación)         │
│  skills/jw-*/SKILL.md   packages/jw-agents/             │
└──────────────────────────────┬───────────────────────────────┘
                               │
┌──────────────────────────────▼───────────────────────────────┐
│  Superficies                                              │
│  • CLI          packages/jw-cli/   (Typer + Rich)      │
│  • Servidor MCP packages/jw-mcp/   (FastMCP)            │
│  • RAG          packages/jw-rag/   (BM25 + vec + RRF)   │
└──────────────────────────────┬───────────────────────────────┘
                               │
┌──────────────────────────────▼───────────────────────────────┐
│  jw-core                                                   │
│  ├─ clients/    cdn · mediator · wol · pub-media · ...   │
│  ├─ parsers/    reference · article · daily-text · ...   │
│  ├─ integrations/ jw-library · obsidian · meps           │
│  ├─ cache · throttle · telemetry · auth · models         │
└──────────────────────────────┬───────────────────────────────┘
                               │
              jw.org · wol.jw.org · b.jw-cdn.org · data.jw-api.org

Long-term vision

Towards a complete local AI ecosystem.

Of the 18 current categories, 13 are already shipped. F65-F76 closed 4 new ones (multimodal·OCR, verifiable reasoning, diachronic analysis, visual verification) and moved Family/kids and Accessibility to partial (F76 consented family voice). The only remaining gap is E2E sync.

Read the full vision
01

Weekly meeting

F57

02

Ministry

F12+F66

03

Audio · TTS · Diarization

F34+F64

04

Multimodal · OCR

F69+F70+F71

05

Languages · ASL

F53+F8

06

Apologetics

F32+F67

07

Operational infra

F9+F10

08

Local privacy

F61+F76

09

Memory · profile

F61.8

10

Family · kids

F76

11

Calendar

F51

12

Accessibility

F68+F76

13

Personalization

F57.16

14

Bible Knowledge Graph

F58+F72

15

Verifiable reasoning

F65+F67

16

Diachronic analysis

F72

17

Visual verification

F70

18

E2E sync

Open source · GPL-3.0

If you build something with AI and public content from jw.org, this is your starting point.

View source code Read the guides