Firecrawl open‑sources web‑agent
Firecrawl open‑sourced a framework called 'web‑agent' that lets developers build agents which search, scrape and interact with web pages using any LLM. The release positions the tool as a 'bring‑your‑own‑model' option for production automation across web workflows. (x.com)
Web agents are software that can search pages, read them, and click through sites like a person using a browser. Firecrawl said on April 16 it open-sourced a framework for building those agents with a developer’s choice of language model. (firecrawl.dev) Firecrawl’s release is called `firecrawl-agent`, and the public GitHub repository was updated within the past day under Firecrawl’s account. The company said the package is a lighter foundation than its hosted `/agent` product, which runs on Firecrawl’s own Spark 1 models. (github.com) (firecrawl.dev) The framework is built around three web tasks: `/search` to find pages, `/scrape` to extract content, and `/interact` to automate browser actions on sites that need clicks or typed input. Firecrawl said developers can scaffold a project with templates for Next.js, Express, or a plain library module. (firecrawl.dev) (github.com) In plain terms, this is infrastructure for teams that want an agent to gather data from websites without handing over the whole stack to one model vendor. Firecrawl said users can swap in Anthropic, OpenAI, Google, or their own model and deploy the code on their own infrastructure, while still using Firecrawl’s web-access API. (firecrawl.dev) That arrives as more software companies split AI systems into two layers: the model that reasons, and the tools that fetch data or act on the web. Firecrawl’s own documentation pitches its service as a way to give agents live web access through search, scraping, crawling, mapping, and browser interaction. (docs.firecrawl.dev) (github.com) The open-source package also shows how Firecrawl is packaging its hosted agent into reusable parts. The README says the stack uses Deep Agents from LangChain for the plan-and-act loop, parallel sub-agent spawning, and `SKILL.md` files for reusable playbooks, while Firecrawl wires in the web tools and structured output. (github.com) Firecrawl has been expanding that ecosystem for months. Its GitHub organization lists a core scraping API with about 110,000 stars, a command-line tool for agent access to web tasks, a Model Context Protocol server, and earlier open-source projects including Open Agent Builder. (github.com 1) (github.com 2) The immediate test is whether developers adopt the open stack instead of only buying the hosted agent. Firecrawl’s pitch is that teams can start with a working web agent in two commands, then change the model, logic, and deployment to fit their own workflows. (firecrawl.dev)
