pageboy

Turn any web article into a queryable MCP server. Point it at a URL, it extracts the readable content, chunks it, indexes it, and exposes search/get tools to any MCP-aware client (Claude Code, Codex, etc.) — all from a single local SQLite file.

• Storage: SQLite + FTS5 (BM25). One file, zero infra.
• Semantic search: optional, via sqlite-vec. Embeddings can be local (Transformers.js or Ollama) or remote (OpenAI).
• Hybrid: combines FTS5 and vector with Reciprocal Rank Fusion.
• Collections: multiple isolated knowledge bases in the same DB.
• MCP server: stdio transport, five tools.

Install

Node ≥ 20.

# From npm (recommended)
npm install -g pageboy

# Or run without installing
npx pageboy serve

From source:

git clone https://github.com/VicenzoMF/pageboy.git
cd pageboy
npm install
npm run build

What you get by default

The default install is intentionally lean — FTS5 (BM25) full-text search works out of the box, with no extra downloads:

Optional: enable semantic search and rerank

Both local embeddings (Transformers.js) and the cross-encoder reranker share one optional dependency:

# Local, fully offline (~50 MB ONNX runtime + models on first use)
npm install @huggingface/transformers

Already have Ollama or an OpenAI key? You can skip the install entirely — pageboy will use them instead (details in Embeddings).

The @huggingface/transformers package is declared as an optional dependency, so it does not install by default with npm install -g pageboy. The package is sizable (it ships an ONNX runtime); we keep it opt-in so the base CLI stays small. Install it whenever you want local embeddings or reranking — pageboy auto-detects it at runtime.

Quick start

# Index an article (FTS only, no embeddings needed)
pageboy add https://example.com/some-post

# List what you've indexed
pageboy list

# Search (defaults to hybrid if embeddings exist, else FTS)
pageboy search "what does it say about hooks"

The DB lives at ./data/pageboy.db (override with --db <path> or PAGEBOY_DB=...).

Wire it into Claude Code

# If installed globally
claude mcp add pageboy pageboy serve

# Or via npx (no install required)
claude mcp add pageboy npx pageboy serve

Or add to .mcp.json in any project:

{
  "mcpServers": {
    "pageboy": {
      "command": "npx",
      "args": ["pageboy", "serve"]
    }
  }
}

Restart Claude Code. Tools available:

CLI commands

pageboy add <url> [-c <collection>] [--no-embed]
                   [-r|--recursive] [--max-depth N] [--max-pages N]
                   [--cross-origin] [--delay-ms N]
                   [--include <regex>] [--exclude <regex>]
pageboy list [-c <collection>]
pageboy collections
pageboy search <query...> [-m fts|vector|hybrid] [-c <collection>] [-n <limit>]
                          [--rerank] [--rerank-model <id>]
pageboy refresh <id|--all> [-c <collection>] [--no-embed]
pageboy embed [-c <collection>]
pageboy doctor           # diagnose DB, providers, reranker, deps
pageboy serve            # start MCP server on stdio

Add --db <path> to any command to target a different SQLite file.

Troubleshooting

Stuck? Run:

pageboy doctor

It prints what's working and what isn't — Node version, DB stats, sqlite-vec load state, which embedding provider is active, whether @huggingface/transformers is installed, and reranker status. Most "it didn't work" reports come down to one missing item in that output.

Recursive crawl

# Index up to 50 pages, depth 2, staying on the same origin
pageboy add https://example.com/docs --recursive --max-depth 2 --max-pages 50

# Limit by URL pattern
pageboy add https://example.com/docs -r \
  --include '^https://example\.com/docs/(api|guides)/' \
  --exclude '\\.(png|jpg)$'

Each page is fetched, extracted with Readability, chunked, and indexed. Same-origin is the default; pass --cross-origin to follow external links.

Reranking

Reranking runs a cross-encoder over the top candidates to reorder them by true relevance to the query. It's optional, and needs the same @huggingface/transformers package as local embeddings:

npm install @huggingface/transformers        # one-time
pageboy search "how does the auth token rotate" --rerank

Default model: Xenova/ms-marco-MiniLM-L-6-v2 (~80 MB, downloaded on first use). Override with --rerank-model <id> or PAGEBOY_RERANK_MODEL=....

For the MCP server, set PAGEBOY_RERANK=1 to rerank on every call.

If @huggingface/transformers is not installed and you pass --rerank, pageboy prints a warning and falls back to the un-reranked results.

Collections

Each doc lives in a collection (default: default). Use collections to keep distinct knowledge bases — say, separate stripe-docs from internal-runbooks — in the same SQLite file:

pageboy add https://stripe.com/docs/api -c stripe-docs
pageboy add https://example.com/runbook -c runbooks
pageboy search "rate limit" -c stripe-docs

Embeddings (optional, but recommended)

Without embeddings you get FTS5 only — fast, but vocabulary-literal: queries must share words with the text. Add embeddings to get semantic + hybrid search that handles paraphrases and cross-language queries.

pageboy picks a provider in this order:

1. OpenAI — if OPENAI_API_KEY is set
2. Ollama — if reachable at http://127.0.0.1:11434 or OLLAMA_HOST is set
3. Transformers.js — fully local, no API. Used when @huggingface/transformers is installed (it is opt-in; install with npm install @huggingface/transformers).

If none of the three is available, pageboy stays on FTS5 only — add and search keep working without embeddings.

Force a specific provider:

export PAGEBOY_EMBED_PROVIDER=ollama       # or openai, transformers
export PAGEBOY_EMBED_MODEL=nomic-embed-text # provider-specific default otherwise

Ollama example

ollama pull nomic-embed-text
pageboy add https://example.com/article          # auto-embeds with Ollama
pageboy search "paraphrased question" -m hybrid

Transformers.js

Install once:

npm install @huggingface/transformers

Default model: Xenova/all-MiniLM-L6-v2 (384 dim, ~25 MB, downloaded on first use). To go lighter or stronger:

PAGEBOY_EMBED_MODEL=Xenova/bge-small-en-v1.5 pageboy embed

Prefer to skip the ~50 MB ONNX runtime? Just don't install @huggingface/transformers — use Ollama or OpenAI instead, or stay on FTS-only search.

OpenAI

export OPENAI_API_KEY=sk-...
# Optional:
export PAGEBOY_EMBED_MODEL=text-embedding-3-small
export PAGEBOY_EMBED_DIM=1536
pageboy add <url>

Retroactive embedding

If you indexed docs before configuring a provider, fill in the missing embeddings later:

pageboy embed                # everything missing
pageboy embed -c stripe-docs # only one collection

Refresh

Articles change. Re-fetch and re-index in place (IDs are preserved):

pageboy refresh 1            # one doc
pageboy refresh --all        # everything

A content-hash check makes no-op refreshes essentially free: if the page hasn't changed, only fetched_at is updated, no chunks are touched.

Architecture

URL ─► fetch ─► JSDOM + Readability ─► section-aware chunk ──┐
                                                              ├─► SQLite (FTS5)
                                                              └─► sqlite-vec (optional)
                                                                     ▲
                                                          embeddings provider
                                                       (OpenAI / Ollama / Transformers.js)

                                  search ─► FTS + vector ─► RRF ─► [cross-encoder rerank]

• Chunking: walks the Readability HTML, tracks H1/H2/H3 path, prefixes each chunk with Title > H2 > H3 so both FTS and embeddings get section context. Soft target ~1400 chars, breaks on paragraph boundaries, sentence-splits long paragraphs.
• FTS5: unicode61 tokenizer with diacritic folding. BM25 ranking.
• Vector store: vec0 virtual table from sqlite-vec. Embeddings stored as raw FLOAT[dim] blobs.
• Hybrid: RRF (k=60) over the top-N FTS hits and top-N vector hits.
• Reranker (optional): Transformers.js cross-encoder (Xenova/ms-marco-MiniLM-L-6-v2 by default) over the top 4× candidates. Toggle with --rerank on the CLI or PAGEBOY_RERANK=1 for the MCP server.
• Recursive crawl: opt-in BFS (--recursive) with --max-depth, --max-pages, same-origin by default, --delay-ms, and --include/--exclude regex filters.
• Migrations: PRAGMA user_version controls schema versions; old DBs are migrated forward on open.

Environment variables

Limitations

• One writer at a time (SQLite WAL). Fine for personal use, not for shared servers.
• Fetches static HTML only (no headless browser). SPAs and pages behind Cloudflare challenges may extract poorly. Pre-rendered docs (Next.js SSG, Stripe, Hugo, etc.) work well.
• Recursive crawl is BFS over anchor tags; it does not parse sitemaps and ignores robots.txt. Use --include/--exclude and --max-pages to keep it polite.
• Embeddings are computed once per chunk; switching providers/models with different dims requires re-embedding (or a fresh DB).
• The cross-encoder reranker downloads ~80 MB on first use and adds ~50–200 ms per query.

License

MIT.