Books Mandala — Find a Book

Purpose

Search, browse, and look up books from Books Mandala — Nepal's largest dedicated online bookstore (50,000+ titles, with deep coverage of Nepali / Hindi / Sanskrit / South-Asian literature and international bestsellers). Returns title, author(s), ISBN, formatted NPR price + numeric value, stock status (in / out), genres, description, and a canonical booksmandala.com/books/{slug}-{id} purchase link for each match. Read-only — never adds to cart, places an order, or hits authenticated endpoints.

When to Use

• A user (anywhere in the world) is looking for a specific Nepali or South-Asian title and wants to know whether Books Mandala stocks it, at what price, and where to buy.
• Discovery flows: "show me bestsellers", "what's new", "books in the nepali-literature genre", "everything by Buddhisagar".
• Agentic shopping where you need a stable, structured catalog surface — Books Mandala publishes a public MCP server and a documented REST API specifically so AI agents can browse the catalog programmatically.
• Anywhere you'd otherwise scrape booksmandala.com HTML — the MCP / API path is ~100× cheaper and structurally more reliable than rendering the Next.js front-end.

Workflow

Books Mandala ships three machine-friendly surfaces; pick in this order of preference:

1. MCP server (recommended, no auth) — https://bm-agent-mcp.booksmandala.workers.dev/mcp. Public, no API key required, 7 tools covering every browse / lookup use case. Streamable HTTP, MCP protocol 2025-03-26.
2. REST API (requires key) — https://booksmandala.com/api/agent/v1. Same 7 operations as the MCP, returns clean JSON, but every endpoint except /health and /docs requires an X-API-Key header. Keys are issued by emailing dev@mandalatech.io. Rate limit 100 req/min per key.
3. Browser fallback — only when MCP is down and you have no API key. The site is Next.js SSR + JSON-LD; works without stealth or proxies.

1. Recommended path — MCP (Streamable HTTP, no auth)

If your agent host (Claude Desktop, Cursor, etc.) supports MCP config, just add it:

{
  "mcpServers": {
    "books-mandala": {
      "url": "https://bm-agent-mcp.booksmandala.workers.dev/mcp"
    }
  }
}

If you're a code-driven agent without a built-in MCP client, talk to it over plain HTTP with two POSTs (initialize + tools/call). The server returns SSE-framed JSON-RPC responses (event: message\ndata: {...}\n\n):

# 1. Initialize → server returns an mcp-session-id header you must echo on subsequent calls
INIT=$(curl -s -D - -X POST https://bm-agent-mcp.booksmandala.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"my-agent","version":"1.0"}}}')
MCP_SID=$(printf '%s' "$INIT" | grep -i '^mcp-session-id:' | awk '{print $2}' | tr -d '\r')

# 2. (optional but spec-compliant) — fire-and-forget initialized notification
curl -s -X POST https://bm-agent-mcp.booksmandala.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "mcp-session-id: $MCP_SID" \
  -d '{"jsonrpc":"2.0","method":"notifications/initialized"}'

# 3. Call a tool — search_books, get_book, list_genres, browse_genre, bestsellers, new_arrivals, get_author
curl -s -X POST https://bm-agent-mcp.booksmandala.workers.dev/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "mcp-session-id: $MCP_SID" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"search_books","arguments":{"query":"palpasa cafe","limit":3}}}'

The 7 MCP tools and when to use each:

Response shape. Every tool returns {"content": [{"type": "text", "text": "..."}]} — a single Markdown-formatted text block, NOT structured JSON book objects. Each book entry looks like:

**Palpasa Cafe**
by Narayan Wagle
Price: NPR 595 | In Stock
ISBN: 9789937905855
Genres: Nepali, Nepali Literature

<full description>

Buy: https://booksmandala.com/palpasa-cafe-2739

If your downstream system needs structured fields, parse with a per-entry block split on \n---\n and a regex pass per line (^by\s+(.+)$, ^Price:\s+(NPR \d[\d,]*)\s*\|\s*(In Stock|Out of Stock)$, ^ISBN:\s+(\d+)$, etc.). For machine-clean JSON, use the REST API instead.

2. Alternative — REST API (key required, returns clean JSON)

Same operations as the MCP but with structured Book objects. Request a key from dev@mandalatech.io, then:

curl -s -H "X-API-Key: $BM_KEY" \
  "https://booksmandala.com/api/agent/v1/search?q=palpasa+cafe&limit=3"

Endpoints (all GET, all under https://booksmandala.com/api/agent/v1):

The OpenAPI spec at /docs defines the canonical Book / BookDetail schemas — fetch it once and use it as your typed contract. Without X-API-Key, every non-/health / non-/docs endpoint returns HTTP 401 {"error":"API key required...","code":"AUTH_REQUIRED"}.

3. Browser fallback (when no MCP + no API key)

Bare session is fine (no Akamai, no proxies, no stealth needed):

sid=$(browse cloud sessions create --keep-alive | jq -r .id)
export BROWSE_SESSION="$sid"

Three browser sub-flows depending on intent:

Search — the homepage has no ?q= URL param. /search?q=... is a 404. Search is a Ctrl+K modal that fires inline typeahead results:

browse open "https://booksmandala.com/" --remote
browse wait load --remote
# Click the "What do you want to read ? Ctrl + K" button (top right of nav)
browse click "button:What do you want to read*" --remote
# Type into the searchbox (don't press Enter — autocomplete renders as you type)
browse fill "searchbox:Search By Title*" "palpasa cafe" --remote
browse wait timeout 2000 --remote
browse snapshot --remote
# urlMap entries shaped like https://booksmandala.com/books/{slug}-{id} are the matches

Genre browse — direct URL works: https://booksmandala.com/books/genres/{genre-slug} (e.g. /books/genres/nepali, /books/genres/fiction-and-literature). Add ?sub_genres={sub} to scope to a sub-genre (e.g. /books/genres/self-improvement-and-relationships?sub_genres=self-help). Genre slugs match list_genres output exactly.

Book detail — https://booksmandala.com/books/{slug}-{id}. Either browse get markdown body to get clean Markdown of the page, or grep the raw HTML for "@type":"Book" — the Next.js hydration payload embeds JSON-LD with the full schema.org Book object (title, image, ISBN, description, publisher, price NPR, availability, aggregateRating). Note the JSON-LD is embedded inside a <script>self.__next_f.push(...)</script> Next.js streaming chunk, not a clean <script type="application/ld+json"> tag — use a substring search for "@type":"Book" then carve from the surrounding object literal.

Release the session when done: browse cloud sessions update "$sid" --status REQUEST_RELEASE.

Site-Specific Gotchas

• The MCP server is genuinely public — don't bother passing API keys. bm-agent-mcp.booksmandala.workers.dev requires no auth header, no client registration, no OAuth. Verified 2026-05-20: a cold initialize from a fresh Browserbase IP returned protocolVersion 2025-03-26 and a session id in the first response. The REST API at /api/agent/v1 is the surface that requires X-API-Key; the MCP is the unauthenticated mirror.
• MCP returns Markdown text, not JSON book objects. Despite being a structured tool-calling protocol, every Books Mandala tool returns a single {"type":"text","text":"..."} block formatted for human/LLM reading. If you need typed fields, either (a) regex-parse the text block, or (b) switch to the REST API which returns proper {"data": Book, "meta": {...}} JSON.
• MCP transport is Streamable HTTP with SSE framing. Responses look like event: message\ndata: <json>\n\n. If your HTTP client treats the body as plain JSON it will fail to parse — strip the event: / data: prefixes first, or use a real MCP client. Also: you MUST send Accept: application/json, text/event-stream (both) on every POST; sending only application/json is rejected.
• MCP session id is REQUIRED on every call after initialize. The server sets mcp-session-id: <hex> in the response headers of the initialize call; echo it back as the mcp-session-id request header on every subsequent tools/call / tools/list. Calls without it return an MCP protocol error.
• REST API rate limit is per endpoint and SHARED with the docs endpoint. Observed: X-Ratelimit-Limit: 20 on the /docs endpoint, and the documented per-key rate is 100/min. Don't loop on /docs — fetch it once, cache it.
• URL paths use /books/{slug} (plural), not /book/{slug} (singular). The llms.txt at https://booksmandala.com/llms.txt documents the singular form; the live site serves the plural. All MCP Buy: links and the front-end's urlMap use plural.
• Slugs carry a numeric id suffix. Canonical URL shape is /books/{kebab-title}-{id} (e.g. /books/palpasa-cafe-2739, /books/atomic-habits-15626). The get_book tool accepts either ISBN-13, the bare slug, or the full slug-with-id.
• Prices are in NPR (Nepalese Rupees) in the API, but the front-end auto-converts based on browser geolocation. Same book on the rendered homepage shows $1.56 while bestsellers MCP returns NPR 209. If you're routing the user to the front-end for purchase, surface both — the user may see a different display currency.
• /search?q=... is a 404 page. The site has no URL-routed search. Browser fallback for search MUST go through the Ctrl+K modal (see workflow). The search button has accessible name "What do you want to read ? Ctrl + K".
• The Ctrl+K modal does NOT navigate on Enter. Results render inline in the modal as a typeahead. Use browse snapshot after fill (with a 1–2s wait for the debounce) and read result anchors from the urlMap. Pressing Enter just submits without going anywhere useful.
• JSON-LD on book pages is embedded inside Next.js hydration chunks, not in clean <script type="application/ld+json"> tags. A naive grep '<script[^>]*ld+json' finds nothing parseable. Search for the literal string "@type":"Book" and carve the enclosing {...} — or just use browse get markdown body, which gives a clean prose version of the same data.
• No anti-bot, no stealth needed. Cloudflare in front of the site (cf-ray header present) but no challenges, no captcha, no IP-based blocking observed. A bare Browserbase session loads the homepage and any deep link cleanly. --proxies and --verified are NOT required for any of the three paths.
• get_book description fields can contain Nepali (Devanagari) script and the MCP / API both preserve UTF-8 correctly. Don't assume ASCII-only output when displaying descriptions.
• Phase 1 only. The MCP and REST API are documented as Phase 1 (discovery / browsing). Real-time stock-by-ISBN, shipping-cost-by-city, and order placement are advertised as Phase 2+ and are NOT in the current tool set. Don't attempt to place an order through any of these surfaces.
• llms.txt and llms-full.txt are the canonical reference. https://booksmandala.com/llms.txt is the short overview; https://booksmandala.com/llms-full.txt is the full catalog + taxonomy + tech doc. Both update; pull them at runtime if you need the genre tree or the latest endpoint list.

Expected Output

When the caller asks "find {query} on Books Mandala", normalize to this shape regardless of which path you used:

{
  "query": "palpasa cafe",
  "method": "mcp",
  "total_results": 3,
  "page": 1,
  "books": [
    {
      "title": "Palpasa Cafe",
      "authors": ["Narayan Wagle"],
      "isbn": "9789937905855",
      "price": "NPR 595",
      "price_value": 595,
      "currency": "NPR",
      "in_stock": true,
      "genres": ["Nepali", "Nepali Literature"],
      "description": "Felicitated by Madan Purashkar in the year 2005…",
      "url": "https://booksmandala.com/books/palpasa-cafe-2739"
    }
  ]
}

For a single-book detail lookup (get_book / /books/{id}):

{
  "method": "mcp",
  "book": {
    "title": "Palpasa Cafe",
    "alternate_title": "पल्पसा क्याफे",
    "authors": ["Narayan Wagle"],
    "isbn": "9789937905855",
    "publisher": "Nepalaya",
    "pages": 286,
    "cover_type": "paper back",
    "edition": "1",
    "weight_grams": 240,
    "languages": ["Nepali"],
    "genres": ["Nepali", "Nepali Literature"],
    "average_rating": 5,
    "reviews_count": 1,
    "price": "NPR 595",
    "price_value": 595,
    "currency": "NPR",
    "in_stock": true,
    "description": "…full text…",
    "url": "https://booksmandala.com/books/palpasa-cafe-2739"
  }
}

Empty / not-found result:

{
  "query": "definitely not a real book xyz",
  "method": "mcp",
  "total_results": 0,
  "books": []
}

Always include the canonical https://booksmandala.com/books/{slug}-{id} URL on every result so the caller has a one-click path to the purchase page. Prices are NPR; surface conversion to the user's locale on the caller side if needed.