Scripture research over MCP
Darash — Bible research your AI can actually check
Ask an AI what the Hebrew says and it answers from memory. Darash gives it the text itself — 59 translations, word-by-word morphology, Strong's numbers, 446,544 cross-references and 13 dictionaries — over the Model Context Protocol, with the source attached to every answer so you can check it.
- 59
- translations, 36 languages
- 446,544
- cross-references
- 31,167
- verses parsed word by word
- 217,352
- dictionary entries
- 14,197
- Hebrew and Greek lexicon entries
- 45
- day free trial
The problem it solves
A language model has read the Bible many times over, which is exactly why it will answer a question about the Greek without looking at any Greek. The answer sounds right. Often it is right. You cannot tell which, and neither can it.
The failure is quiet. A Greek verb in the active voice renders naturally as a passive in English, and an argument about who is acting on whom, built on that English passive, is an argument about the translator rather than about the text. Nothing in the answer tells you that happened.
Darash removes the guess. Your assistant asks for the parsing and gets voice, person and stem as the text carries them; it asks whether a lemma is in a verse and gets a yes or a no rather than something remembered from a concordance. Every answer arrives with its source, so the claim lands back on the text where you can check it.
Is this the one you want?
We make five of these and they do different jobs. Only this page has a reason to tell you when the answer is one of the others.
| If you want to | Use |
|---|---|
| Read the Hebrew or Greek, and check it | Darash — this one |
| Write a book, study or sermon series and publish it | Junifye |
| Keep the terms you define, with sources and versions | Lexifye |
| Look up a Norwegian company or its accounts | Brreg |
| Read the current text of Norwegian law | Lexar |
| Convert a currency at a published central-bank rate | Currency |
| Just read a Bible, on a phone, offline | BibleRead |
Who it is for
- Someone working in the biblical languages who wants the parsing and the lexicon, not a paraphrase.
- A congregation, college or ministry that needs it run privately. Darash also ships as a single binary of about 205 MB carrying CLI, HTTP API and MCP server, working entirely offline — for an institution whose material cannot leave the building, that is usually the only workable shape.
- A developer putting Scripture data into their own product. Every tool, its input schema and the provenance of every dataset are published before you hold a credential.
Questions people ask first
What is Darash?
Darash is a Scripture research engine from Publifye AS, a publishing house in Oslo. It speaks the Model Context Protocol, so the AI assistant you already use — Claude, Cursor, Claude Code or your own agent — reads the primary texts alongside you: 59 translations in 36 languages including the Leningrad Codex and the Byzantine Greek, Strong's numbers, word-by-word morphology, 446,544 cross-references and 13 reference works.
Do I need to install anything?
No, and that is the point. Darash is an MCP server, so it plugs into the AI you already use. You paste one URL and sign in from the browser. You keep your assistant, your habits and your notes, and it gains the scriptures.
Is this a Bible reader?
No. A reader hands you someone's conclusion. Darash hands you the text and the data underneath it — the Hebrew and Greek, the parsing, the lexicons, the cross-references — so you can see whether those things are so. If you want something to read on a phone, BibleRead is the right product and this is not.
Can I try it before paying?
Yes — a 45-day free trial, 1,500 calls a day, no card. That is the real engine and not a demo corpus: the same translations, the same lexicons, the same cross-references. Current prices are on the Darash site.
Can we run it on our own servers?
Yes. Self-hosted puts the binary inside your own network, so queries from your students or staff never leave it. To be exact about what does leave: the instance checks its licence and periodically reports an aggregate call count, a number, for licence administration. No query, reference, search term or result is ever sent.
How do I know the data is what it claims to be?
Every dictionary and both Strong's lexicons carry a provenance record: the printed work, the edition, how the text reached us, the licence, and what is NOT established about it. One is marked verified, meaning it was collated against the printed edition it claims to be; the rest are marked conventional, meaning they were not. We once served a dictionary under the name Thayer that was not Thayer, found it, and rebuilt the lexicon from the 1889 page scans — 716 pages, two passes agreeing at 98.91% across 856,793 tokens. A corpus that claims perfection has stopped looking.