Pali Canon for Claude

The Theravāda Buddhist scriptures, word for word, with exact citations: for Claude and the age of AI.

Pali Canon for Claude gives Claude the complete Pāli Tipiṭaka, the scripture of Theravāda Buddhism. With it, Claude can read the actual text, quote it accurately, cite it precisely, and show published English translations side by side. It runs on your own computer and the Pāli text works without internet.

Version 2.1.0 · Free and open source (MIT) · An independent community project, not affiliated with Anthropic or any Buddhist publisher.

Contents

• Why this exists

• What it does

• What you can ask it

• What's inside

• Our approach

• Sources at a glance

• How accuracy is checked

• Install

• Using it well

• Typing Pāli on an English keyboard

• What it does not do yet

• Roadmap

• Documentation

• Contributing, privacy and licence

Related MCP server: tripitaka-mcp

Why this exists

Theravāda Buddhism is the living tradition of Sri Lanka, Myanmar, Thailand, Cambodia and Laos, with well over 100 million followers worldwide. Its scripture, the Pāli Canon (Tipiṭaka, "three baskets"), is one of the largest scriptures of any religion: about 2.8 million words of Pāli.

Our mission is to bring the rigour of traditional and academic Buddhist scholarship together with the best openly available Pāli and Theravāda resources, in one modern tool built for AI. We started with Claude. The same tool now also works in Codex, Gemini CLI and code editors, with ChatGPT and a hosted web version next. All the while we are building a proven, verifiable repository of knowledge: every passage comes from a named source and every claim can be checked.

When this project started, we found no Theravāda tool in Claude's connector directory. Meanwhile, AI models answering questions about the Buddha's teaching often:

• paraphrase from memory and present it as a quotation;

• invent or garble citations ("MN 152" for a passage that is not there);

• confuse the different numbering systems that translators use;

• repeat popular quotes that are not in the texts at all.

What it does

Once installed, you can ask Claude things like:

Claude quotes only what it retrieved, names the translator, and keeps four things separate: the Pāli text, the translation, the dictionary meaning, and its own explanation.

What you can ask it

Pali Canon gives the AI 32 tools, 4 prompts, 3 reference resources and 5 skills. You do not need to know their names: ask in your own words and the AI picks the right tool. The tool names are shown in code for anyone who wants to check. Everything works offline unless marked (internet).

Study

• Find and read any text. Give a citation, a PTS page or a link and get the Pāli with its references (resolve_reference, get_metadata, get_text). Browse the canon book by book (browse_canon) or search the whole Pāli text (search_texts).

• Read the Pāli in your own script. Sinhala, Thai, Myanmar, Devanagari, Khmer, Lao or Lanna (Tai Tham), transliterated from the VRI edition; you can also type citations and search words in those scripts (script on get_text, search_texts and resolve_reference). Lao and Lanna have no VRI script edition to test against yet.

• Cite the Thai, Myanmar or VRI printed edition. "Thai MN 1.103", "Myanmar Dhp 15" or "VRI SN 1.20 (Nidānavaggo)" finds the text on that page, and get_text marks where each edition's pages begin. See citation formats.

• See where the printed editions differ. VRI's variant notes, split into readings with the editions (Sinhala, Thai, PTS and others) that carry them (get_text with variants: "structured").

• Read and compare translations (internet). Ṭhānissaro Bhikkhu, the translators on Access to Insight and Ānandajoti Bhikkhu, always named and attributed (get_translation, compare_translations, search_access_to_insight, get_access_to_insight_text).

• Look up a Pāli word in the Digital Pāḷi Dictionary (lookup_pali, live, answering from a bundled subset when offline), or analyse a passage word by word: split, grammar, root and meaning of every word (analyse_passage). A pronunciation guide explains how to read Pāli aloud (pali://pronunciation).

• Ask in English or type Pāli without diacritics. "Find the Satipatthana sutta", "lookup nibbana" or "search mindfulness of breathing" work: names, plain spellings and English words are understood, and the answer always says what was inferred (glossary, and every search and lookup tool). See Typing Pāli on an English keyboard.

• Read the commentary. The traditional Pāli commentary (Aṭṭhakathā) on a text, aligned by VRI paragraph and always labelled as commentary, not canon (get_commentary, or get_text with commentary: true).

• Find parallels and cross-references. Other texts that share a passage with this one word for word, and texts the traditional commentary cites when explaining it, each link saying where it comes from and how it was matched, plus the SuttaCentral link for parallels in other traditions (get_parallels). See D59.

• Read later Pāli classics. The Visuddhimagga ("Vism 1.1", "Vism §19"), the Abhidhammatthasaṅgaha ("Abhidh-s 1.4"), the sub-commentaries (Ṭīkā) and chronicles such as the Mahāvaṃsa, labelled canonical: false (get_post_canonical, search_post_canonical, or search_texts with include_post_canonical). Visuddhimagga PTS pages cannot be looked up yet.

• Ask who or where. "Who was Anāthapiṇḍika?" lists the texts that name a person or place and links to the Dictionary of Pāli Proper Names (lookup_name).

• Explore by topic or list. "Texts on anger", "What are the five hindrances?": over 400 topics from Access to Insight's indexes and 37 numbered lists with their source passages (topic, list_topics, dhamma_list). See topics and lists.

• Abhidhamma and Vinaya. The matrix (mātikā) of the Dhammasaṅgaṇī (abhidhamma_matika), the 24 conditions of the Paṭṭhāna (patthana_condition), and any of the 227 monks' rules with a link to the Buddhist Monastic Code that explains it (vinaya_rule).

• Find Dhamma talks (internet) by Ṭhānissaro Bhikkhu on a topic or a sutta, as links only (find_talks).

• Study skills. study reads and explains a passage, compare compares translations, pali-word explains a word in context, and method keeps genre, early text, commentary and school differences apart.

Practice

• Uposatha and festivals. Observance days and the dates of Vesak, Āsāḷha Pūjā, the Rains Retreat and more, for Thai Mahānikāya, Thai Dhammayut, Myanmar and Sri Lanka (uposatha, festivals). The dates are computed and labelled as such; your monastery's calendar decides. See the calendar guide.

• Chanting. Paritta and other commonly recited texts (Maṅgala, Ratana, Karaṇīyamettā and more) in Pāli, with links to English (paritta).

• Daily reading. A Dhammapada verse or a sutta for the day (daily_text), and reading plans such as the Dhammapada in 26 days or the Majjhima Nikāya in a year (reading_plan).

• Learn by heart. The memorise skill (and the memorise_passage prompt in Claude Desktop) breaks a passage into short parts, quizzes you with fill-in-the-blanks and suggests when to review.

Checking quotes and citations

• "Did the Buddha really say this?" verify_quote checks a quotation. Pāli is checked against the whole canon offline; English against the linked translations (internet), by default the Dhammapada. It says whether the words were found, closely paraphrased or not found in the texts searched, and it never claims who did or did not say something. See checking a quotation.

• Check every citation in an essay. Paste an essay, a talk transcript or a reference list, and find_citations finds each citation (MN 10, S v 420, links…) and checks that it points to a real text.

Every tool only reads; none can change anything. The full technical list is in docs/development.md, and more example requests are in docs/examples.md.

What's inside

• The Pāli is complete, from the Vipassana Research Institute's Chaṭṭha Saṅgāyana edition (the text of the Sixth Council, Yangon, 1954–56): 61 books, bundled and available offline.

• Page references for the PTS edition are included: 11,964 page markers, so scholarly citations such as "M i 55" work. Thai, Myanmar and VRI page markers (56,050) are shown and can be cited too.

• English translations by Ṭhānissaro Bhikkhu (dhammatalks.org), the many translators on Access to Insight and Ānandajoti Bhikkhu (Ancient Buddhist Texts) are fetched live, with backup copies if a site is down.

• A Pāli dictionary: the Digital Pāḷi Dictionary, live, with a bundled subset for word-by-word analysis and offline lookups.

• The commentaries (Aṭṭhakathā) aligned to the texts, and 109 post-canonical books (the Visuddhimagga, Abhidhammatthasaṅgaha, sub-commentaries and chronicles), from the same VRI edition, bundled and always labelled as not canon.

The full breakdown by collection is in docs/coverage.md.

Our approach

1. Retrieve, never recall. Claude reads the text from the bundled edition or a named translation every time. Nothing is generated from memory.

2. Cite precisely. Every passage carries its standard citation (MN 10), the VRI paragraph, the PTS volume and page, and the translator.

3. Keep the layers distinct. Scripture, translation, dictionary gloss, and Claude's own interpretation are always labelled separately.

4. Be honest about uncertainty. Different translators number some texts differently, and the canon abbreviates repeated passages. The tool reports both rather than hiding them.

5. Honour the wishes of sources. We use sources whose terms allow this use, credit them, and do not use material from people who have asked that their work not be used in AI tools.

6. Prefer what lasts. The Pāli is bundled, so it keeps working offline and if websites disappear. Online sources have official mirrors.

7. Prove it. Every one of the 8,544 texts is checked automatically before each release (see below).

The reasoning behind each choice is recorded in docs/decisions.md.

Sources at a glance

Considered and not used (with reasons), and the sources planned next, are listed in the source library.

How accuracy is checked

• Every text, every release. An automatic audit checks all 8,544 texts: each resolves by its citation, has text, does not overlap its neighbours, and its PTS page leads back to it. Result: 0 failures.

• Checked against an independent edition. All 1,190 canonical pages of Access to Insight and their PTS citations were compared with our mapping: 0 problems.

• Translations are linked only with strong evidence. A translation is attached only when its title, number or PTS page matches unambiguously. We prefer no link to a wrong link. 2,540 links pass this test (64 of them placed by comparing the names in the English with the Pāli, each audited).

• Independent review. A separate reviewer checked about 110 citations and every PTS lookup in its sample. The problems it found were fixed and are now permanent tests.

• Outage drills. We simulate each website going down and confirm that the backups take over and the Pāli still works.

Details: docs/how-it-works.md.

Evaluation

We measure whether answers cite correctly instead of just saying so. The suite in evals/ has 189 questions: the kind of thing a user asks, the citations a correct answer must give, and checks that call the server's tools. Categories: citation and PTS-page resolution, numbering traps (AN 7.64 vs AN 7.68; the Kālāma Sutta is AN 3.65 here but AN 3:66 on dhammatalks.org), abbreviated texts, quote checking, edition pages, scripts, variants, the calendar, lists and topics, Vinaya rules, Abhidhamma, reading plans, paritta, translation availability and refusals of invalid citations. Every expected citation is checked against the bundled canon each time the suite runs.

Offline tool checks (npm run evals, run by CI on every change), November 2026:

Model-level runs, where Claude or Codex answers the questions with the server connected and must cite the expected texts, are scripted in evals/claude-plugin-eval/ (claude plugin eval) and evals/codex/ (codex exec). They need a model and are run by hand; their scores will be added here once run.

Install

Pick the app you use. Every option runs the same server, with the same 32 tools.

Claude Desktop (easiest)

Full guide: docs/install/claude-desktop.md.

1. Download the latest pali-canon-<version>.mcpb file from the Releases page.

2. Open Claude Desktop → Settings → Extensions, and drag the file onto the page (or use Advanced settings → Install Extension…).

3. Click Install and make sure the extension is switched on.

4. Ask: "Use Pali Canon to read MN 10 with a translation and cite the PTS page."

No account, key or extra software is needed. Claude Desktop provides everything the extension needs.

Claude Code

Install Node.js 20 or newer, then type in Claude Code:

/plugin marketplace add DaniilKi/pali-canon-claude
/plugin install pali-canon@pali-canon

Start a new session and try the five built-in skills:

/pali-canon:study MN 10
/pali-canon:compare AN 3.65
/pali-canon:pali-word anicca in SN 22.59
/pali-canon:method Kv 1
/pali-canon:memorise Dhp 1

Full guide, updating and troubleshooting: docs/install/claude-code.md.

The skills also work in other AI apps that read SKILL.md files (npx skills add DaniilKi/pali-canon-claude), but they need the Pali Canon server installed too.

Codex (CLI and app)

Pali Canon works in OpenAI's Codex as a plugin (tools plus the $study, $compare, $pali-word, $method and $memorise skills) or as a plain MCP server (tools only). You need Node.js 20 or newer.

# Plugin: add this repository as a marketplace, then run /plugins inside Codex and install "Pali Canon"
codex plugin marketplace add DaniilKi/pali-canon-claude

# Or just the server, from a downloaded copy of this repository
codex mcp add pali-canon -- node /absolute/path/pali-canon-claude/server/index.cjs

Full steps and troubleshooting: docs/install/codex.md. This setup has not yet been tried in a real Codex install; please tell us how it goes.

Gemini CLI and Antigravity CLI

Install Node.js 20 or newer, then run:

gemini extensions install https://github.com/DaniilKi/pali-canon-claude

Antigravity CLI users can then import it with agy plugin import gemini. Full steps: docs/install/gemini.md.

Code editors and agent apps

Cursor, VS Code, Windsurf, Zed, Cline and Goose: step-by-step setup (needs Node.js 20 or newer) is in docs/install/ide.md.

Note. The buttons above, and every npx -y pali-canon-mcp command, will work once the npm package is published. That has not happened yet. Until then, download this repository and use node with the full path to the server instead: node /path/to/pali-canon-claude/server/index.cjs. The editor guide shows how.

Local AI models

Running a model on your own computer (LM Studio, Open WebUI, Ollama)? See Use it with local AI models. The Pāli text, search and citation lookup work fully offline.

Hosted version and Docker

A web version lets Claude on the web and on phones connect by address, with nothing to install. It is not online yet: the code is ready, but it has not been deployed to a hosting service. This section will give the address once it is live.

You can already run your own copy with Docker or Node.js; see docs/hosting.md. It needs no sign-in and logs no request content.

Other MCP apps

Technical users can connect any app that supports MCP (the Model Context Protocol). Use the command npx with the arguments -y pali-canon-mcp once the npm package is published, or today node /path/to/pali-canon-claude/server/index.cjs. See docs/development.md.

ChatGPT

The hosted version can be added to ChatGPT as an app, with a reading view that shows the Pāli and English side by side, with the citation, PTS page and translator. It is not submitted yet, because it needs the hosted version first. See docs/listings/chatgpt.md.

Using it well

• Cite the way you normally would. MN 10, SN 56:11, AN 3.65, Dhp 183, Snp 1.8, Thag 1.1, Ja 1, Pj 1, Kd 1, PTS pages like M i 55, or a link from SuttaCentral, Access to Insight or dhammatalks.org all work. Every accepted form is listed in the citation guide.

• Ask for a translator by name, for example "with Ṭhānissaro's translation", or ask to compare all of them.

• When the tool reports a numbering difference, keep both numbers in what you write.

• Remember it is a library, not a teacher. It gives you the texts and their translations faithfully. Interpretation, and questions of practice, belong with the tradition and its teachers.

New to the terms? See the glossary.

Typing Pāli on an English keyboard

You do not need Pāli letters. Type the way your keyboard lets you, and every answer says what it understood, so nothing is silently changed.

• No diacritics: "samyutta", "nibbana", "satipatthana". Search ignores diacritics anyway. A dictionary lookup lists every word spelled with those letters: "nibbana" gives both nibbāna (extinguishment) and nibbana (desireless), marked inferred_spelling.

• Common English-keyboard spellings: "sanyutta" (saṃyutta), "nyana" (ñāṇa), "nirvana" (nibbāna, marked as a Sanskrit or English form). When an exact search finds little, the search also tries spellings that differ only in h after a consonant, double letters, ny/ñ or ṃ/n, and says so (spelling_note).

• Typing schemes: Velthuis ("sa.myutta", "pannaa", "nibbaana") and Harvard-Kyoto ("saMyutta"). Velthuis dots and tildes never occur in English words, so they are always read. Harvard-Kyoto capitals, doubled vowels and English-keyboard spellings are read only when the result is a word in the canon, so ordinary English ("vacuum", "McDonald", "NASA") is left alone.

• Sutta names: "Satipatthana Sutta", "Dhammacakkappavattana", "anapanasati sutta" or an English title such as "Mindfulness of Breathing" find the text (matched_by: "name"). When a name belongs to several texts, as the Metta Sutta does (Snp 1.8, Kp 9 and others), you get the list to choose from; the tool never picks one for you. "samyutta 56 11" is read as SN 56.11.

• Questions in English: a search such as "mindfulness of breathing" or "suffering" is turned into Pāli words (sati, assāsapassāsa; dukkha…) from the Digital Pāḷi Dictionary's English index, and the answer lists exactly which Pāli words were searched (expanded_from_english). Set english: "off" to search only what you typed. The glossary tool looks words up in both directions. Glossary meanings are dictionary labels for finding words, not translations of the texts.

What it does not do yet

• English is linked for 1,823 texts, about a fifth of the canon. The Vinaya, nearly all of the Abhidhamma and much of the Khuddaka Nikāya have no English yet.

• The commentaries (Aṭṭhakathā) and sub-commentaries (Ṭīkā) are not included yet.

• Other scripts are transliterated by this server from the Roman edition, not taken from printed script editions. Lao and Lanna have not yet been checked against a printed edition or by native readers.

• Translations and the dictionary need internet; the Pāli does not.

• It runs on your own computer. A hosted version for Claude on the web and phone is built but not online yet.

Roadmap

Fifty issues are planned, under one tracking issue: Roadmap: parity with Bible and Quran AI tools (#52). In short:

• Reach: putting the hosted version online, Claude's connector directory, ChatGPT, the MCP Registry and community directories, and support in many languages. The Codex, Gemini CLI and code editor packages are done and waiting for hands-on testing.

• Study: commentaries, word-by-word analysis, parallels between texts, more translators and other languages. Quote checking, citation finding, topics, numbered lists, people and places, other scripts, edition page citations and structured variants are done. Search by meaning was tested and deferred (why).

• Practice: the uposatha calendar, festivals, chanting texts, daily readings and memorisation are done. Chanting-book layouts and recordings wait on publishers' permission.

How we compare with Bible, Quran and Sefaria tools, and why each item matters: docs/roadmap.md.

Documentation

The docs site shows these pages in a browser (it goes live once GitHub Pages is switched on). AI agents can read llms.txt.

Contributing, privacy and licence

• Found a wrong citation or a broken link? Please open an issue with the citation you used, what you expected (ideally with a PTS page) and what you got. Translators, scholars, monastics and native speakers of Sinhala, Thai, Burmese, Khmer and Lao are especially welcome. See CONTRIBUTING.md.

• Privacy: there are no accounts, analytics or stored questions. The Pāli is read on your computer. Translation and dictionary requests go to those public sites or their mirrors, which may keep ordinary web logs. See the privacy policy and terms of use.

• Security problems: please report them privately; see SECURITY.md. Everyone taking part follows our code of conduct.

• Licence: the software is MIT. The texts are not: each keeps its own terms, which are mostly non-commercial. Do not sell content retrieved with this tool. See THIRD_PARTY.md.

Acknowledgements

With gratitude to the Vipassana Research Institute; Ṭhānissaro Bhikkhu and Metta Forest Monastery; the Barre Center for Buddhist Studies and John Bullitt for Access to Insight; Bodhirasa Bhikkhu and the Digital Pāḷi Dictionary team; the Pali Text Society; and SuttaCentral, whose numbering we follow and whose wishes we respect.

Sabbadānaṃ dhammadānaṃ jināti: the gift of the Dhamma surpasses all gifts (Dhp 354).