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)
sefaria_text	Text of any reference (Hebrew + translation): 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 an MCP prompt (/mcp__torah-mcp__hebrewbooks).

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-Agent and 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.