firecrawl / web-agent

Firecrawl open-sourced the same agent it runs at firecrawl.dev/app/agent: an orchestrator called agent-core, built on LangChain's Deep Agents, wired to Firecrawl's own search, scrape and interact (browser automation) tools. The six skills sit inside it as auto-discovered playbooks — competitor-analysis builds a comparison matrix across pricing, features and positioning; deep-research triangulates three or more sources for a research question; e-commerce pulls products, pricing and variants from a storefront, pagination and all; financial-research combines SEC EDGAR filings with Yahoo Finance and TipRanks for a public company; pricing-tracker normalizes a vendor's pricing tiers and pairs with an export step for a recurring cron job; and structured-extraction pulls data matching a JSON schema out of a page. All six are tagged "ready" in our dependency check, meaning nothing beyond the harness itself is required to load them.

This is for a developer who wants a working example of an autonomous web-research agent to fork and extend, not someone looking to add one skill to a Claude or Copilot setup they already have. The repo ships full Next.js and Express deployment templates around agent-core, and the quick start is scaffolding a project (firecrawl create agent -t next), not installing a skill file.

READ THE FULL ANALYSIS＋

What running it actually needs. The readme's own quick start is npx firecrawl-cli@latest init -y --browser, which installs and authenticates the Firecrawl CLI before anything else works — the six skills call Firecrawl's search/scrape/interact tools, which are billed API calls against a Firecrawl account, not local logic. Bring your own model key too: agent-core is the orchestrator, not the model provider.

Read from Firecrawl's own repository capture; the agent-core skill definitions themselves were not opened individually, only their frontmatter descriptions.