Pith

The essential web, for agents. URL → clean markdown, crawl, structured extract, and search — with verifiable citations and deterministic request replay.

Pith is an open-source web-for-LLMs engine: point it at a URL and get clean, LLM-ready markdown — plus multi-page crawl, provider-agnostic structured extraction (with per-field confidence and citations you can verify), web search, and an MCP server. It runs with zero infrastructure and zero API keys by default: in-memory stores, an in-process queue, no database, no object store, no job runner. Postgres / MinIO / BullMQ / LLM providers are optional adapters behind explicit ports.

It's the core of a production web-for-LLMs platform, carved out as a library you can embed or self-host.

Install

npm install @use-pith/core
# optional peers, installed only if you use that face:
npm install fastify               # for @use-pith/core/http
npm install @modelcontextprotocol/sdk zod-to-json-schema  # for @use-pith/core/mcp
npm install playwright            # for the headless fetch tier

Related MCP server: WebScrape MCP Server

Quickstart (SDK, zero env)

import { createEngine } from "@use-pith/core";

const pith = createEngine();          // in-memory defaults — no keys, no containers

// Scrape: static → headless escalation, boilerplate stripped before Readability.
const page = await pith.scrape("https://example.com");
console.log(page.markdown);

// Crawl: same-domain bounded, dedup, worker-death resumable.
const handle = await pith.crawl("https://example.com", { maxDepth: 2, maxPages: 25 });
const status = await handle.wait();   // drains to a terminal status

// Structured extract + verifiable citations (needs a backend — see below).
const backend = createEngine({
  extractionBackend: createExtractionBackend({
    baseUrl: process.env.EXTRACTION_BASE_URL!,
    apiKey: process.env.EXTRACTION_API_KEY!,
    model: process.env.EXTRACTION_MODEL!,
  }),
});
const { data, confidence, citations, flaggedFields } = await backend.extract(
  "https://example.com",
  { type: "object", properties: { title: { type: "string" } }, required: ["title"] },
);

createEngine() with no arguments works immediately. extract and search throw NotConfiguredError until you pass a backend — Pith never assumes a billable provider.

The three faces

// HTTP
import { createEngine } from "@use-pith/core";
import { createServer } from "@use-pith/core/http";
const app = await createServer({ engine: createEngine() });
await app.listen({ port: 3000, host: "127.0.0.1" });

// MCP
import { buildMcpServer } from "@use-pith/core/mcp";
const server = await buildMcpServer({ engine: createEngine() });
// …connect the SDK's StreamableHTTPServerTransport (see examples/mcp/)

Use it from any MCP client (e.g. Claude Desktop)

buildMcpServer({ engine }) returns a standard MCP server — connect it to any transport. The simplest is a small stdio bridge (examples/mcp/server.ts):

// examples/mcp/claude_desktop_config.json  (run from the repo root, or adjust the path)
{
  "mcpServers": {
    "pith": { "command": "npx", "args": ["tsx", "./examples/mcp/server.ts"] }
  }
}

The 5 tools (scrape, search, crawl, get_crawl_status, extract) then appear in the client. Full setup in examples/mcp/.

Configuration

Everything is optional. Pith ships no default billable endpoint.

See .env.example for the full list (zero required).

Why Pith

• Defense-in-depth content cleaning — boilerplate is stripped before Readability runs, so the reader sees real article text, not nav chrome.

• Provider-agnostic, verifiable extraction — every field carries a confidence score (FR-4) and a citation; citations are independently re-verified against the page's own text (FR-8), so hallucinated quotes are caught, not just claimed against. Works against any OpenAI-compatible endpoint.

• Tier escalation — static fetch first, headless only when the result is thin or the static fetch fails; a budget ceiling can decline an escalation that has a cheaper fallback.

• Deterministic request inspection / replay — every request is snapshot-able and replayable.

• Opt-in freshness — stale-while-revalidate scrape cache with a monotonically-tightening tier model.

• Test-first quality gate — a 5-project Vitest matrix (unit / smoke / integration-real / integration-nock / accuracy) stays green key-free on every PR; the 20-fixture extraction accuracy benchmark runs nightly (key-gated), hard-asserting ≥90% field accuracy and ≥90% citation verification against real model output.

Architecture

Pith is one core, three faces. Every host concern (cost metering, request snapshots, crawl state, content blobs, the job queue, robots resolution, the freshness cache, the clock) is an injectable port with an in-memory / no-op default. Swap in real adapters without touching engine logic.

                 ┌──────────── createEngine() ────────────┐
   SDK ─────────▶│ scrape · crawl · extract · search      │
   HTTP ────────▶│ (one request-handling core, shared)    │◀─▶ CorePorts (8 ports, null defaults)
   MCP  ────────▶│                                        │     └ real adapters optional (PG/MinIO/BullMQ/…)
                 └────────────────────────────────────────┘

Read docs/ARCHITECTURE.md for the port table, the "no infrastructure on import" invariant, and the carve-out boundary.

Examples

• examples/quickstart.ts — Node/TS SDK (scrape, crawl, extract, search).

• examples/langchain_agent.py — a LangChain agent driving Pith's MCP tools (search → scrape → extract).

• examples/mcp/ — MCP client config (Claude Desktop).

Contributing

See CONTRIBUTING.md. Quick rules enforced in CI: no infrastructure on import, test-first (the matrix stays green key-free), no secrets / no real crawled HTML. Security issues: see SECURITY.md — do not open a public issue.

License

Apache-2.0 — see LICENSE.