Torah MCP
A source discipline for Claude — never quote Torah from model memory. Before answering any religious question, Claude loads a bundled study method that forces it to read the actual texts (Sefaria API), quote exactly, flag divergent opinions, and point to hebrewbooks.org for further study — never fabricating a reference. A HebrewBooks catalog search (~65,000 seforim) completes the loop.
Pairs ideally with Sefaria's official MCP
(https://mcp.sefaria.org/sse) — install both: the official server for deep
library access (14 tools), Torah MCP for the citation discipline, the study
method and HebrewBooks.
Free, no account, no data collected. Hosted instance:
https://torah-mcp.com/mcp
Install
claude.ai — Settings → Connectors → Add custom connector → paste the URL above. Works on mobile too once added.
Claude Code
claude mcp add --transport http torah https://torah-mcp.com/mcp
Any MCP client — streamable HTTP transport at the same URL.
Tools
| Tool | Purpose |
|---|---|
hebrewbooks_skill | The core: the study method loaded before answering any religious question — sources actually read, exact quotes, hebrewbooks.org reading links |
hebrewbooks_search | HebrewBooks catalog search (~65,000 scanned seforim) by title/author via their official API — requires a HEBREWBOOKS_API_KEY (granted on request: developers@hebrewbooks.org) |
mode_etude | Study register — debutant (all in French, every term explained, for readers with no background or no Hebrew), classique (bilingual, default), avance (beit midrash: original text, no glosses, mahloket, girsaot, lomdus). Same source discipline in all three |
havrouta_mode | Study-partner mode: Claude asks the questions, makes you defend Rashi against Tosafot, never hands out answers |
guide_paracha | AlHaTorah-style weekly study guide: aliyot, three textual questions with diverging commentators, haftarah echo, Shabbat-table questions |
daf_viewer | MCP App: an interactive Vilna page in the conversation (Gemara, Rashi, Tosafot, click-to-translate). Also at /daf |
zmanim · date_hebraique · gematria · nikoud · fiche_source | Daily-life tools: Hebcal times, Hebrew dates, 5-method gematria, Dicta vocalization, WhatsApp-ready source cards |
sefaria_text | Text of any reference (Hebrew + translation — French first when Sefaria has it, e.g. Bible du Rabbinat for Tanakh, else English): Berakhot 2a, Shulchan Arukh, Orach Chayim 1:1… |
sefaria_links | Linked commentaries and sources (Rashi, Tosafot, midrash, halakhah…), filterable by category |
sefaria_search | Full-text search across the whole library (Hebrew or English) |
sefaria_calendar | Parashat hashavua, haftarah, daf yomi, daily Rambam… |
The server also ships an instructions block (loaded by MCP clients at
initialize) telling the model to load the skill before answering any religious
question, and five MCP prompts (hebrewbooks, havrouta, paracha, debutant, avance).
The website — Mamash IA
The website now lives under its own brand, Mamash IA (the connector keeps
its torah-mcp identity; torah-mcp.com redirects, and torah-mcp.com/mcp keeps
serving installed connectors forever). Everything the MCP does also exists on the web, no install needed, in three
languages: French at the root (mamash-ia.com),
English under /en, Hebrew (RTL) under
/he — same slugs everywhere (/question, /daf,
/outils, /install, /daily, /privacy), a FR · EN · עב switcher on
every page, hreflang links for search engines.
/question— ask in plain language; Claude runs server-side with the same tools and method and answers in the language of the page, with the sources it actually read (copy / share / local history)./daf— the Vilna-style daf viewer;/outils— zmanim, Hebrew date, gematria, nikkud, source sheets;/daily— today's study cycles.
Optional secrets for the site: ANTHROPIC_API_KEY (enables /question),
STATS_PASSWORD (private /stats page over the D1 question log — no IPs).
Deploy your own
Or manually:
npm install
npx wrangler deploy
No secrets required. Optional: set a BEARER_TOKENS secret (comma-separated
list) to switch from public access to invitation-only — each guest then uses
https://<your-worker>/<token>/mcp, and you revoke by removing a token.
Good-citizen notes
- Requests to Sefaria carry an identifying
User-Agentand are edge-cached (24 h for texts and links, 1 h for calendars) so repeated study of the same daf costs Sefaria a single API call. - A per-IP rate limit (60 req/min) protects the public endpoint.
- Text licenses are returned with every response. Data by the wonderful Sefaria — support them.
License
MIT — see LICENSE. The bundled study-method skill and this server are unaffiliated with Sefaria and hebrewbooks.org.