Web crawler MCPs (3-tier fallback)

runbook mcpcrawlerscraperfirecrawlapifysmitheryrunbook

Web crawler MCPs (3-tier fallback)

Why

URL → crawl N pages → markdown for AI agent ingestion. The three tiers degrade gracefully: best quality first, then a service fallback, then a no-key local crawler that always works.

Per the dual-bucket MCP rule: keyed services land in Smithery (token never enters the public repo); the no-key local crawler lands in .mcp.json (safe to commit).

Tier 1: firecrawl (Smithery, KEYED — best quality)

JS-rendered scrape + structured crawl with markdown output. Highest quality for SPAs and gated pages.

1. Signup: https://firecrawl.dev/signup — free 500 credits/mo, no card.

2. Copy FIRECRAWL_API_KEY from the dashboard.

3. Install via Smithery (paste the key on prompt):

   npx -y @smithery/cli@latest install @mendableai/firecrawl-mcp-server --client claude
   

Tools exposed: firecrawl_scrape, firecrawl_crawl, firecrawl_map, firecrawl_search.

Tier 2: apify (Smithery, KEYED — fallback)

Massive actor catalog (Web Scraper, Cheerio Scraper, Puppeteer, etc.) reachable through one MCP. Use when firecrawl is over quota or a site needs a specific actor (e.g. Instagram, Twitter, Maps).

1. Signup: https://console.apify.com/sign-up — $5 free credit/mo, no card.

2. Copy APIFY_TOKEN from Settings → Integrations.

3. Install via Smithery (paste the token on prompt):

   npx -y @smithery/cli@latest install apify/actors-mcp-server --client claude
   

Tools exposed: dynamic — depends on which Apify actors you authorise in the Smithery profile.

Tier 3: crawl-md (in-repo, NO KEY — last resort)

Local Node.js MCP using fetch + turndown. No external service, no key, no quota. Slower (no JS rendering, single-threaded BFS), but always works and never costs anything.

• Source: scripts/crawl-mcp/server.js
• Auto-loaded via .mcp.json → no install step needed once the repo is cloned.
• Tool: crawl(url, max_pages=50, same_origin=true) → markdown of all crawled pages.
• Hard cap: 200 pages per call. HTML only (skips PDF/images/binaries).

Fallback decision flow

1. Try firecrawl_crawl — best quality, handles JS.
2. If 402/quota or auth-walled site fails → try apify actor (e.g. cheerio-scraper for static HTML, puppeteer-scraper for JS).
3. If both keyed services are out of credits → fall back to crawl-md for static HTML.

Related

• keyed-mcp-via-smithery-2026-06-27 — Smithery install pattern, key rotation.
• Decision: dual-bucket MCP rule (no-key in .mcp.json, keyed in Smithery).