SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformstripitaka-mcp

tripitaka-mcp

Search, cite and compare the Pāli Canon (Tipiṭaka): hybrid search, full suttas with cross-references, aligned translations and Pāli dictionary lookup.

Open website

About this tool

Purpose

tripitaka-mcp is an MCP (Model Context Protocol) server for retrieving and citing content from the Pāli Tipiṭaka. It gives AI agents (such as Claude or Cursor) the ability to find suttas, cite teachings, and compare translations across languages. The project is offered as "Dhamma Dāna" (the gift of Dhamma), 100% free, and restricted to non-commercial use.

Data Coverage

Aligned with SuttaCentral, it includes the entire Tipiṭaka (approximately 444,000 segments):

  • Sutta Piṭaka: Pāli + Sujato English
  • Vinaya Piṭaka: Pāli + Brahmali English
  • Abhidhamma Piṭaka: Pāli only (there is no English for any Abhidhamma book in the upstream bilara-data)

Main Capabilities

  • Hybrid Search: combines keyword and semantic search, achieving higher precision via Reciprocal Rank Fusion (RRF).
  • Keyword search: trigram fuzzy matching, supporting cross-language alignment.
  • Semantic search: semantic retrieval based on vector similarity (pgvector).
  • Translation comparison: view and compare translations across different versions, aligned at the segment level.
  • Dictionary bridging: built in with over 20,000 entries (P. A. Payutto, PTS, DPPN).
  • Retrieve suttas and citations: retrieve sutta content by ID (such as mn1, pli-tv-bu-vb-pj1, patthana1.1) and generate properly formatted academic citations.
  • Pāli word-form analyzer: when a dictionary lookup fails, strip inflectional suffixes to find the root (e.g., bhikkhūnaṁ → bhikkhu).
  • Cross-reference URLs: each response includes a clickable deep link to the project's own bilingual reader (Pāli + English, with segment anchors to highlight the cited verses). The reader renders SuttaCentral's bilara-data verbatim, so it is the authoritative text.
  • Dual transport modes: supports both traditional SSE (/sse) and the canonical Streamable HTTP (/mcp, MCP specification 2025-03-26).
  • MCP resources: tripitaka://structure, tripitaka://sutta/{id}, tripitaka://word/{w}, for clients that pin context as resources.
  • Curated reference pages (/topics/*): six markdown pages covering the structure of the canon, getting started and choosing tools, places (the sixteen great nations + sacred sites + cosmology), 10 foundational topics with canonical sources, about 30 major figures, and a timeline of the Buddha's 45-year teaching career divided into phases. Sutta IDs have been verified against live data.
  • Claude skill: skills/tipitaka-research.md provides a ready-to-install workflow file that activates a multi-step research mode (clarify → verify coverage → search → go deeper → cite) on Claude Desktop / Claude Code.
  • Postman support: comes with a Postman collection for testing the API.

How to Use

No setup: connect to the public Dhamma Dāna server

The maintainers run a free public instance at tripitaka-mcp.com. Connecting Claude Desktop only requires Node.js—no installation, Docker, or GPU:

  1. Find the absolute path to npx (Claude Desktop does not read shell configuration files, so a bare npx will not resolve).
  2. Open claude_desktop_config.json (on macOS at ~/Library/Application Support/Claude/, on Windows at %APPDATA%\Claude/) and add the corresponding entry, replacing YOUR_NPX_PATH with the output from step 1 and YOUR_NODE_BIN_DIR with the parent directory of that path.
  3. Fully quit Claude Desktop and reopen it. The 🔌 indicator in the lower-left corner should show tripitaka and offer 12 tools.

The first connection takes 5–10 seconds, because npx needs to download mcp-remote on demand. The hosted server has rate limits (10 requests/10 seconds + 60 requests/minute per IP).

Run fully offline (pipx — local SQLite, no server)

The local version places the entire PāliThe Tipitaka is packaged as a single SQLite file (about 120 MB) and runs as a local stdio MCP server. If pipx uses an unexpected interpreter, you can explicitly specify the version (3.10 or higher is fine). After that, just point Claude Desktop / Cursor to the local command; no npx, mcp-remote, or internet is needed. If you need a URL instead of stdio, you can bind to 127.0.0.1 (unless otherwise specified) to provide an HTTP endpoint.

Quick Local Installation (Recommended for Non-Developers)

The installer downloads a prepared database dump from Hugging Face (dhamma-seeker/tripitaka-mcp-dump) and restores it automatically, reducing setup time from 2–4 hours to about 5 minutes. The installer will: verify whether docker, compose, openssl, and curl are installed; generate a .env with a random password; download the dump; start the database and restore it; set up a read-only role and runtime timeouts; and print a Claude Desktop configuration that can be pasted directly.

Developer Manual Setup and Production Deployment

The production stack runs 3 services: db (PostgreSQL + pgvector, internal only), mcp-server (FastMCP, running as a read-only user), and caddy (reverse proxy + Let's Encrypt + rate limiting). See DEPLOYMENT.md for details.

Differences Between Hosted and Local Versions

Both serve the same Tipitaka of about 444,000 segments. The differences are: search_semantic / search_hybrid and the trigram keyword index require PostgreSQL + pgvector + an approximately 1 GB embedding model, which is too heavy for a lightweight local installation, so they are provided only in the hosted version. In local mode these two tools are not registered, and connecting clients will only see 9 available tools. The local server is a standard stdio MCP server and can be paired with local models (such as Ollama) and any MCP-compatible chat interface to achieve a fully offline AI stack.

Limitations and Notes

  • The vector index for search_semantic is built only on text_pali (SuttaCentral's bilara-data does not yet have a Thai translation), using an unspecializedmultilingual MiniLM model trained for Pali. Therefore Pali/English queries are more accurate, while Thai queries match loosely and are not recommended. For exact keywords such as appamāda, search_by_keyword is more precise; for general searches, search_hybrid has the best tolerance. There are plans to upgrade to a Pali-trained embedding model (such as bge-m3) and embed the Thai version.
  • The total number of MCP tools is 13.
  • License: The project aggregates data from multiple sources with different licenses; please read NOTICE.md in full before redistributing. It is allowed for free/dhamma-giving/educational projects and personal use; it is prohibited for any paid product or service (due to the Payutto dictionary); modifying the dictionary content is prohibited. For commercial use, you must remove the dictionary component or contact Wat Nyanavesakavan for permission.
  • The hosted server is intended for personal study, research, and Dhamma practice; please refer to NOTICE.md before redistribution or commercial use.

Related tools

Browse category