Darash

Darash — Bible research MCP server for Claude, Cursor and any MCP client

Darash gives an AI assistant the Hebrew and Greek text of Scripture as structured data: which stem a verb is in, which Strong's numbers actually occur in a verse, what a named 19th-century lexicon says about a lemma, and which printed edition that lexicon came from. It runs as a hosted MCP server over HTTPS — there is nothing to install.

Endpoint	https://darash-api.publifye.com/mcp
Transport	Streamable HTTP
Auth	OAuth 2.1 + PKCE (S256), DCR open · or X-API-Key
Registry	pro.publifye/darash (server.json)
Product site	https://darash.publifye.com
What it solves	https://publifye.com/darash
Capability tools	34 (full schemas)

Connect

{ "mcpServers": { "darash": { "type": "http", "url": "https://darash-api.publifye.com/mcp" } } }

The OAuth flow runs in the browser on first use. Full detail, including the API-key route and the discovery chain: ../../docs/connect.md.

The corpus

Measured on the live service, 2026-09-13, via get_health:

Translations	59, across 36 languages
Verses indexed	2,142,531
Cross-references	446,544
Hebrew lexicon entries	8,674
Greek lexicon entries	5,523
Dictionary entries	217,352, across 13 dictionaries
Morphologically parsed verses	31,167
Koren Torah letters	304,805

Included among the 59: the Leningrad Codex (wlc) and the Byzantine Majority Greek text. Strong's is augmented with the full Abbott-Smith, LSJ and BDB.

Why a translation is not enough

A Greek verb in the active voice can be rendered perfectly naturally in the passive in English. An argument about who is acting on whom, built on that English passive, is an argument about the translator rather than about the text.

get_morphology returns the voice, person and stem the text actually carries. strongs_in_verse proves a lemma is genuinely in a verse rather than merely adjacent to it in a concordance. That distinction is the whole reason this server exists.

What the tools do

Every tool is documented with its exact description, annotations and input schema — 42 in all (34 capability, 8 session and figure tools), generated from the service's own tools/list, never written by hand.

Area	The question it answers	Tools
Reading the text	Which verse, chapter or book does the text actually contain, and how do translations differ?	11
Hebrew and Greek	What is this word, grammatically and lexically — voice, person, stem, lemma, and what the lexicons say?	9
Study and cross-reference	What else does Scripture say about this, and which words move together?	6
Dictionaries	What do the reference works say, and which work is it?	4
Search	Find it by wording, by meaning, or by how rare it is.	4
Session, figures and health	How do I check the service, send feedback, and manage the figures I have stored?	8

Machine-readable: tools.json carries all 42 callable tools (34 capability, 8 session/cache) with full JSON Schema, plus every excluded bucket listed by name so the count is auditable.

Provenance — read this before you trust a gloss

PROVENANCE.md records, for every dictionary and both Strong's lexicons: the printed work, the edition, how the text reached us, the licence, and what is not established.

One dictionary is marked verified — collated against the printed edition it claims to be. The rest are marked conventional, meaning the title and author are as distributed and have not been collated. Most vendors in this space ship the same public-domain texts and do not draw that distinction. We draw it because a reader deciding how much weight a gloss can carry needs it.

The file also carries a corrections section. In September 2026 we found that a dictionary served under the key thayer was not Thayer; it was archived with a digest and replaced with a transcription of the 1889 Corrected Edition read from page images — 716 body pages, two independent passes collated at 98.91% agreement over 856,793 tokens, with ambiguous headwords refusing to resolve rather than guessing. The provenance work is what found the error. That is the argument for publishing it.

Self-hosting

Darash also ships as a single self-contained binary of roughly 205 MB, carrying CLI, HTTP API and MCP server together, running fully offline with no network call at any point. For institutions whose material cannot leave the building, that is usually the only workable shape. Licensed by quote — see the product site.

The ELS / Torah-code research tools are part of the self-hosted Engine only. They are not in the hosted plans, are not listed in tools.json, and are enforced server-side rather than merely hidden.

Plans

Three ways in — a personal key for any MCP client, an always-on agent, and a self-hosted or managed licence by quote. There is a free trial. Current plans and prices are on https://darash.publifye.com, which is the only place they are authoritative; they are deliberately not duplicated here, because a price in a repository is a price that goes stale.

Written about this work

Published, each also in Norwegian, Spanish, Chinese and Korean:

• Thayer's Greek Lexicon, rebuilt from the 1889 page scans — 716 pages, four transcription routes, five rounds of adversarial review, and why every open digital Thayer we could find is lossy.
• Darash vs Logos, Accordance and Blue Letter Bible (2026) — an honest comparison — including where the big platforms win.
• Deuteronomy 4:29 says "seek" twice. The Hebrew does not. — a worked example, including the searches that prove nothing.
• Why do AI models hallucinate — and does grounding them in real sources fix it?
• Best Bible study tools in order: what to use first, and when Hebrew and Greek pay off
• Bibelen Anno 2026 — a Norwegian translation built with AI, eight verse gates and a human veto