# Firecrawl MCP Server: Search Scrape And Crawl The Web > [!question] Lingering Questions > - How often does an agent silently fall back to default web tools in real sessions, absent a deny rule? > - What's the token and latency cost of Firecrawl's hosted remote server versus local default fetch? > - Does keyless mode carry rate or feature limits that change the comparison? > [!tip] Article Ideas > - A quickstart showing the keyless Firecrawl MCP install plus the CLAUDE.md fallback snippet so agents actually prefer it over defaults. > - A short lab comparing default web fetch vs Firecrawl on one JavaScript-heavy page to make the tooling-vs-model distinction concrete. > [!link] Cross Reference > - [[3. Resources/Permanent Notes/Tool-Interface Choice Shapes Agent Architecture More Than Model Choice]] — same diagnosis: agent failures trace to tooling/interface, not model capability > - [[3. Resources/Permanent Notes/The Harness Now Outpaces the Model as the Source of Agent Capability Gains]] — parallel claim about where agent capability gains actually come from > [!quote] Vocabulary > - *MCP (Model Context Protocol)* — the standard interface letting agents call external tools like Firecrawl > - *harness* — the agent runtime (Claude Code, Codex) that decides which tools an agent can reach for ## Bib capsule - **Source**: Firecrawl-produced YouTube tutorial, ~6 minutes - **URL**: https://www.youtube.com/watch?v=aovuyqN2QPE - **Category**: Vendor product demo / developer tooling ## Author's thesis (Adler Q1+Q2) Default agent web-search and fetch tools return raw HTML and URL lists that the model must parse itself, which fails on JavaScript-heavy pages and produces guesses users wrongly blame on the model; the Firecrawl MCP server fixes this by combining search and fetch into one step that returns clean markdown (and structured JSON), rendering "like a real browser." ## Key arguments (≤3) 1. **The failure is tooling, not model** — raw-HTML fetch + JS-heavy pages forces the model to guess from training data; Firecrawl instead returns rendered, clean content. 2. **Agents don't reliably self-select the better tool** — even after installing Firecrawl, the demo shows the agent falling back to defaults, requiring deny rules or CLAUDE.md config to force adoption. 3. **Structured/location-aware extraction changes what's possible** — location-correct pricing and JSON-only extraction fix categories of error default tools can't fix at all (currency mis-conversion). ## Claim atoms ### Default agent web tools hand back raw HTML, not answers > [!quote] ID:rw_manual > "The agent still has to open the URL and go through the raw HTML just to read it." Context: ~00:40, framing the core problem. Matters because agent web-search failures are often a tooling limitation, not a model limitation — a reframe worth keeping whenever an agent gives a confidently wrong web answer. ### Firecrawl combines search and fetch into clean markdown > [!quote] ID:rw_manual > "it searches the web like a real browser and hands back the full page content in clean markdown" Context: ~01:10, the central product claim — and the design pattern worth generalizing: render like a browser, return structured/clean output in one step. ### Location-aware scraping fixes currency, not just formatting Context: ~03:30 demo. Default tools scraped a Shopify site in dollars and mis-converted to pounds (~170 vs actual 118); Firecrawl scrapes as if browsing from that country and returns the correct local price. This is a correctness bug, not a style preference. ## Permanent candidates - **Agent failures often trace to tooling, not the model** — supported directly by H1/H2, generalizes past this one video into a diagnostic habit for any agent web-research failure. Promoted below. ## Processing checklist - [ ] Watch the linked Firecrawl CLI follow-up video for token-saving patterns - [ ] Test keyless Firecrawl MCP install on Claude Code with a deny rule for default web tools - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] ## Links - [[3. Resources/MOCs/AI & the Future of Knowledge Work MOC]] — the tooling-vs-model diagnostic bears directly on how agent capability is actually delivered - [[3. Resources/Readwise/Full Document Contents/Articles/Firecrawl MCP Server Search Scrape and Crawl the Web]] — source video and full deep-read notes