Firecrawl

Recently shipped a custom relevance model for /search that achieves 94.7% accuracy on SimpleQA.

Most accurate /search yet — A custom model scores every paragraph, list, and table against the query, returning the most relevant excerpts. Agents using it score 94.7% on SimpleQA — higher than any other provider — while using 10x fewer tokens than processing full pages. The API shape is unchanged, so existing calls get better results automatically.

Research Index targets AI/ML research loops. Search 3M+ arXiv papers and GitHub artifacts (issues, merged PRs, READMEs) refreshed daily. 53.3% recall on arXivQA beats the next best provider by 18%. Agents retrieve papers, verify claims against full text, and pull code in one loop. Works with Codex, Claude Code, and any agent harness.

Web-scale /monitor watches the entire web, not just pages. Define search queries and a plain-English goal, and Firecrawl searches on your schedule, alerting via webhook or email when new results match. The same endpoint handles search, deduplication, filtering, and alerts — replaces the DIY stack. A separate /monitor endpoint watches individual pages for changes, with up to 90% fewer tokens by only ingesting what changes.

Scraping improvements across formats and ingestion. The highlights format returns only matching sentences, code blocks, and table rows (up to 100x fewer tokens). question returns grounded answers with prompt-injection isolation, skipping the scrape-parse-prompt pipeline. deterministicJson returns structured JSON without an LLM per request — generates reusable extractors per schema and site. /parse uploads PDFs, Word docs, and spreadsheets up to 50 MB via a Rust engine (up to 5x faster), returning clean Markdown, JSON, or summaries. lockdown: true on /scrape serves results exclusively from Firecrawl's index with zero outbound requests and zero data retention. The video format discovers videos on any page, returning URL, title, description, and thumbnail.

Keyless access and PII redaction. /scrape, /search, /interact, and /parse work without an API key from official MCP, CLI, and SDK clients. The new redactPII option strips names, emails, phone numbers, addresses, and secrets from scraped content.