buddha

Fast Buddhist text search and retrieval for humans and AI agents. This repository provides one Rust CLI (`buddha`) and an MCP stdio server (`buddha mcp`) over local and online Buddhist text corpora.

Languages: English | 日本語 | 繁體中文

> Former project/binary names (`daizo`, `daizo-cli`, `daizo-mcp`) are kept as compatibility aliases, but current development uses the `buddha` name.

What It Covers

Core strengths:

• Direct ID fetch is the fastest path when the text ID is known.
• Search results include line anchors and `_meta.fetchSuggestions` for low-token follow-up fetches.
• CBETA search normalizes common old/new CJK variants so modern forms still find Taisho text.
• CLI JSON output uses MCP-style envelopes for easy parsing by agents.
• MCP defaults to a compact unified tool surface while keeping legacy corpus-specific tools available.
• Large MCP text responses can be spilled to `cache/mcp-spill/` instead of overloading clients.

Install

Prerequisites: Git and Rust/Cargo. The quick installer checks these and suggests Rust installation when needed.

curl -fsSL https://raw.githubusercontent.com/sinryo/buddha-cli/main/scripts/bootstrap.sh | bash -s -- --yes --write-path

Manual build and install:

cargo build --release -p buddha
scripts/install.sh --prefix "$HOME/.buddha" --write-path

The installer places binaries in `$BUDDHA_DIR/bin`, creates `buddha-mcp` plus legacy `daizo*` aliases, downloads or updates local corpora where supported, and rebuilds indexes.

MCP Setup

Run the server as:

buddha mcp

Claude Code:

claude mcp add buddha "$HOME/.buddha/bin/buddha" mcp

Codex (`~/.codex/config.toml`):

[mcp_servers.buddha]
command = "/Users/you/.buddha/bin/buddha"
args = ["mcp"]

Compatibility: `$HOME/.buddha/bin/buddha-mcp`, `daizo`, `daizo-cli`, and `daizo-mcp` all point at the same CLI for older client configs.

MCP Tools

Unified MCP tools are enabled by default (`BUDDHA_UNIFIED_TOOLS=0` disables them):

`info` and `profile` remain callable diagnostics but are hidden from the default model tool list to save schema tokens and avoid accidental benchmarking. Set `BUDDHA_MCP_DIAGNOSTIC_TOOLS=1` to list them. Tool definitions use bounded JSON Schemas and read-only annotations; results keep text in `content` and expose status, cursors, citations, errors, and executable `nextActions` in compact `structuredContent`.

Legacy corpus-specific tools are still available when unified mode is disabled or when an older client has cached tool names, for example `cbeta_fetch`, `cbeta_search`, `gretil_pipeline`, `sat_detail`, `jozen_fetch`, and `buddha_version`.

CLI Quick Start

Use `buddha doctor --json` for a machine-readable inventory of installed binaries and all local corpus paths; each item has an explicit `available` boolean.

Direct fetch when you already know the ID:

buddha cbeta-fetch --id T0262 --max-chars 4000 --json
buddha tipitaka-fetch --id DN1 --max-chars 2000 --json
buddha gretil-fetch --id saddharmapuNDarIka --max-chars 4000 --json
buddha sarit-fetch --id asvaghosa-buddhacarita --max-chars 4000 --json
buddha muktabodha-fetch --id "" --max-chars 4000 --json

Find an ID when you only know a title or alias:

buddha resolve --query "法華経" --json
buddha cbeta-title-search --query "楞伽經" --json
buddha tipitaka-title-search --query "dn 1" --json
buddha gretil-title-search --query "vajracchedika" --json

Search content and fetch context:

buddha cbeta-search --query "阿弥陀" --max-results 10 --json
buddha cbeta-fetch --id T0858 --line-number 342 --context-before 2 --context-after 6 --highlight "阿弥陀" --json
buddha tipitaka-search --query "nibbana|vipassana" --max-results 15 --json
buddha gretil-search --query "dharma" --max-results 10 --json
buddha shinzen-search --query "法身偈" --max-results 10 --json
buddha shinzen-fetch --id SZ09p0243 --line-number 19 --context-before 2 --context-after 4 --json
buddha shinzen-work-catalog --query "即身成佛" --json
buddha shinzen-fetch --id SZW0001 --json
buddha shoku-shinzen-work-catalog --query "" --json
buddha shoku-shinzen-fetch --id SKZW0001 --max-chars 2000 --json

Fetching a `SZW` work ID (from `shinzen-work-catalog`) resolves to the work's first page automatically (`_meta.resolvedFromWork`). SHINZEN JSON responses include best-effort `biblio` metadata for search hits and fetch results. It derives work title/author from public catalog records such as CiNii and NTU Buddhist Digital Library, plus local global TOC images, opening titles, running titles, and colophons, not from the TEI header. Fetch results also include `work` metadata from the precomputed `shinzen-work-catalog.json` and a ready-to-quote `_meta.preferredCitation` (for example 頼瑜『即身成佛義顯得鈔』、真言宗全書 第13巻 047頁). Use `confidence` to decide whether the generated `citation` is citable; catalog entries carry `medium`, `range-only`, or `metadata-only`, where `metadata-only` means the work is citable (title/author verified) but its page range is unresolved, so search within SHINZEN instead of fetching by work ID. Unresolved hits keep page-level citation and a warning.

The SHINZEN work catalog prefers the bundled public-catalog TSV. If you need auxiliary OCR for the SHINZEN global table-of-contents images, use the local `ndlkotenocr-rs` pipeline rather than Tesseract:

scripts/ocr-shinzen-somokuji.sh
buddha shinzen-work-catalog-index

To refresh the precomputed work summaries after improving ranges or metadata:

scripts/build-shinzen-work-summaries.py
buddha shinzen-work-catalog-index

Online sources:

buddha sat-search --query "般若" --json
buddha sat-fetch --useid "" --max-chars 3000 --json
buddha jozen-search --query "念仏" --json
buddha jozen-fetch --lineno "J01_0200B19" --json
buddha tibetan-search --query "bde ba" --json
buddha tibetan-fetch --source adarsha --kdb degetengyur --sutra D3134 --page 74-299b --json

SAT `startid` values are detail `useid`s for fetching the containing text/fascicle, not match-level line anchors. `sat-fetch --start-char` slices the extracted detail text after fetching; it is not an offset into `sat-search` `body`.

Tibetan fetching is backend-dependent: Adarsha can return page text from `kdb + sutra/voltext + page`; BUDA/BDRC tries BDRC e-text chunks first, falls back to BDRC snippets, then RDF/OpenPecha metadata for restricted or scan-only records.

Admin and discovery:

buddha init
buddha doctor --verbose
buddha index-rebuild --source all
buddha schema
buddha schema --command cbeta-fetch
buddha version

Agent Usage Pattern

For low-token AI use:

1. If an ID is known, call `fetch` directly.

2. If the corpus or ID is unclear, call `resolve`.

3. Otherwise call `search`, read `_meta.fetchSuggestions`, then call `fetch` with the suggested `id` plus `lineNumber` or `lb`.

4. Include `highlight` with the original search term when fetching context.

5. Use `pipeline` only when you want a multi-file summary or automated search-to-fetch flow.

Common direct IDs:

> Note: only GRETIL ids carrying the `sa_` prefix auto-detect reliably. Bare file stems for GRETIL/SARIT/MUKTABODHA (e.g. `saddharmapuNDarIka`, `asvaghosa-buddhacarita`) are auto-probed against the local corpora when `source` is omitted; specify `source` explicitly for reliability.

Output And Errors

`--json` returns compact machine-readable JSON. In non-TTY contexts, output is automatically JSON unless overridden.

buddha cbeta-title-search --query "般若" | jq .
BUDDHA_JSON=1 buddha cbeta-title-search --query "般若"
buddha --json cbeta-title-search --query "般若"

JSON errors are written to stderr as:

{"error":{"message":"...","code":"NOT_FOUND"}}

MCP tool failures instead return `result.isError: true` with machine-readable `structuredContent.error` / `_meta.error` (`category`, `code`, `field`, `retryable`) and, when safe, executable `nextActions`. Do not retry unchanged arguments when `retryable` is false. When it is true, retry at most once after a short delay; if that fails, stop and report.

Exit codes:

Directories And Environment

`BUDDHA_DIR` defaults to `~/.buddha`. Legacy `DAIZO_DIR` and `~/.daizo` are still recognized as fallbacks.

Typical layout:

$BUDDHA_DIR/
  bin/
  cache/
  xml-p5/
  tipitaka-xml/romn/
  GRETIL/
  SARIT-corpus/
  MUKTABODHA/
  SHINZEN/kunten_tei_xml/

Useful environment variables:

Most variables also accept legacy `DAIZO_*` names.

Repository Layout

Development

cargo fmt
cargo test -p buddha-core
cargo test -p buddha-mcp
cargo test -p buddha
cargo build --release -p buddha

Golden regression check:

bash tasks/golden/verify.sh local

Golden output includes the CLI version, so a release bump can intentionally change `tasks/golden/*/version.text.out`.

Release

Current release version: `0.6.16`.

Version numbers live in:

• `buddha-core/Cargo.toml`
• `buddha-cli/Cargo.toml`
• `buddha-mcp/Cargo.toml`
• `Cargo.lock`
• `CHANGELOG.md`
• `docs/buddha_system_prompt.txt`

Release helper:

# Inspect the next patch release without modifying files
scripts/release.sh --patch --dry-run --no-fmt --no-test

# Create commit, push branch, create v0.6.15 tag, push tag, create GitHub Release with generated notes
scripts/release.sh 0.6.16 --push --tag --release --auto-notes

# Use CHANGELOG notes instead of GitHub generated notes
scripts/release.sh 0.6.16 --push --tag --release

Important: `scripts/release.sh` runs `git add -A` and commits. Review the working tree before using it, especially when unrelated changes are present.

License

MIT OR Apache-2.0 © 2026 Shinryo Taniguchi