Firecrawl Skills

Firecrawl Skills is the official Agent Skills catalog from Firecrawl — 33 skills that teach AI coding agents to search, scrape, crawl, and interact with the live web through the Firecrawl CLI, integrate the Firecrawl APIs into product code, and run end-to-end research and lead-generation recipes.

What is Firecrawl?

Firecrawl turns web pages into clean markdown built for LLM context windows. It handles JS-rendered SPAs, whole-site crawls, page interaction (clicks, forms, pagination, login), scheduled change monitoring, and local file parsing for PDF, DOCX, and XLSX. It runs as a hosted API with SDKs for Node, Python, Rust, Java, and Elixir, as a CLI, and as an MCP server — or self-hosted.

Three categories, 33 skills

Core (12) — the primitives for live web work in a session, driving the Firecrawl CLI: firecrawl (the umbrella router), firecrawl-search, firecrawl-scrape, firecrawl-crawl, firecrawl-map, firecrawl-interact, firecrawl-agent for AI-powered structured extraction, firecrawl-monitor for scheduled change detection, firecrawl-parse for local files, and firecrawl-download. Two index skills back dedicated commands: firecrawl-research-index for biomedical and scientific literature across PubMed, bioRxiv, medRxiv, and arXiv, and firecrawl-developer-index for answering library, API, and bug questions from issues, merged PRs, READMEs, and docs.

Build (5) — integrating Firecrawl into your own codebase: firecrawl-build (the umbrella, with references on endpoint selection, integration patterns, auth, and verification), firecrawl-build-onboarding (getting FIRECRAWL_API_KEY into a project and picking the right SDK), plus firecrawl-build-scrape, firecrawl-build-search, and firecrawl-build-interact for the individual endpoints.

Workflows (16) — end-to-end session recipes: firecrawl-deep-research, firecrawl-lead-gen, firecrawl-lead-research, firecrawl-competitive-intel, firecrawl-market-research, firecrawl-seo-audit, firecrawl-knowledge-base, firecrawl-knowledge-ingest, firecrawl-company-directories, firecrawl-research-papers, firecrawl-dashboard-reporting, firecrawl-website-design-clone, firecrawl-qa, firecrawl-shop, firecrawl-demo-walkthrough, and the firecrawl-workflows umbrella with an authoring guide for writing your own.

The escalation pattern

The umbrella skill teaches a specific order rather than reaching for the heaviest tool first: search when there is no URL yet, scrape when there is one, map + scrape to locate a subpage on a large site, crawl for a whole section, monitor for recurring checks instead of repeated one-off scrapes, and interact only when content needs clicks, forms, or pagination. It also tells the agent to reuse content already fetched — search --scrape returns full pages, and prior results sit in .firecrawl/.

Security posture

Rare among scraping tooling, the catalog ships an explicit prompt-injection policy. Fetched web content is treated as untrusted third-party data: commands write to .firecrawl/ files instead of dumping pages into the context window, the agent is told to read those files incrementally with grep or head rather than whole, .firecrawl/ is gitignored so scraped content is never committed, fetching is user-initiated only, URLs are always quoted against command injection, and instructions found inside fetched pages are explicitly not to be followed.

Who it is for

Developers and researchers who want an agent that can actually read the web — gathering sources, extracting structured data, watching pages for changes — and teams adding Firecrawl scrape, search, or interact calls to a production codebase.

What you can build

Research briefs with cited sources, lead lists built from company directories, competitive intelligence sweeps, SEO audits, knowledge bases ingested from documentation sites, change-monitored dashboards, and scraping integrations in Node, Python, Rust, Java, or Elixir.

Why it matters

Web access is where agents most often fail quietly: a JS-rendered page returns an empty shell, a paginated list stops at page one, a scraped page silently carries injected instructions into the context. This catalog is maintained by Firecrawl itself, routes each task to the narrowest capable command, and treats everything it fetches as hostile input by default.

Option 1 — skills CLI

Interactive picker over the whole catalog:

npx skills add firecrawl/skills

Install one skill by name — names are path-independent, so you do not need the category folder:

npx skills add firecrawl/skills --skill firecrawl-build

Option 2 — Firecrawl CLI onboarding (recommended for live web work)

This installs the CLI, authenticates through the browser, and installs the core, build, and workflow skills across every detected coding editor:

npx -y firecrawl-cli@latest init -y --browser

Safe to re-run when the CLI is missing, stale, or partially configured. To install skills separately once the CLI exists:

firecrawl setup skills
firecrawl setup workflows

Option 3 — Editor plugin

The repository ships plugin manifests for three editors — .claude-plugin/, .cursor-plugin/, and .codex-plugin/. In Claude Code:

/plugin marketplace add firecrawl/skills
/plugin install firecrawl

Note that the plugin bundles the 17 core and build skills only; the 16 workflow recipes are session material and install through one of the paths above.

Authenticate

export FIRECRAWL_API_KEY="your-key-here"

Get a key at firecrawl.dev/app, or let firecrawl-build-onboarding walk the browser authorization flow. Check state at any time:

firecrawl --status

That reports auth state, concurrency limit, and remaining credits.

Optional — MCP server

The plugin bundles configuration for the official Firecrawl MCP server, so editors that read bundled MCP metadata can wire the tools up from FIRECRAWL_API_KEY:

{ "mcpServers": { "firecrawl": { "command": "npx", "args": ["-y", "firecrawl-mcp"],
  "env": { "FIRECRAWL_API_KEY": "${FIRECRAWL_API_KEY}" } } } }

Use it

Describe the web task in plain language — the umbrella skill routes it:

Find the top 20 Series A fintech companies in Europe, pull their pricing pages, and build me a comparison table.

The agent picks search, map, scrape, or crawl in that order of narrowness, writes results to .firecrawl/, and cites the saved source files.