Firecrawl is an open-source, self-developed web proxy framework that supports multiple models and parallel sub-agents, enabling one-click generation of survey tools.

According to ME News, on April 17th (UTC+8), Beating, a monitoring platform, reported that Firecrawl officially open-sourced its Web Agent framework, designed to help developers build AI agents capable of autonomously executing closed-loop tasks of "search-crawl-interaction." Based on Firecrawl's existing /agent architecture, the framework supports integration with Anthropic, OpenAI, and various self-hosted large models, emphasizing flexibility and self-hosting capabilities in web research scenarios. The core architecture employs a "Plan-Act" loop mechanism. After receiving instructions, the agent breaks down the steps and generates parallel "subagents" that work synchronously in independent browser sessions. This design offers significant advantages when handling large-scale concurrent tasks; for example, in the video demonstration, the agent can simultaneously extract real-time stock data and news from multiple company websites. For developers, this tool significantly lowers the barrier to entry for web agent development. The `firecrawl create agent` command can quickly generate complete agent templates based on Next.js or Express. In addition, the framework introduces reusable "skill manuals" (SKILL.md), allowing developers to encapsulate complex workflows (such as specific crawling logic for Yahoo Finance) into skills for the agent to directly invoke in subsequent tasks. In the latest version, the Web Agent has upgraded its default driver model to Claude Opus 4.7 and includes built-in specialized skill sets for the financial sector. The project is currently open-source under the MIT license, encouraging the community to fork and further develop it. (Source: ME)