Open web‑agent framework 'Firecrawl'
Firecrawl published an open‑source web‑agent framework that supports search, scraping and interaction flows and is model‑agnostic, enabling agents to use any underlying model for web tasks. (x.com) The project frames itself as a general web‑automation layer for agents rather than a single‑model solution. (x.com)
Web agents are software that search pages, read them, and click through forms like a person using a browser. Firecrawl said on April 16 it open-sourced a framework for building those agents with search, scraping, and interaction tools in one stack. (firecrawl.dev) Firecrawl’s new project is called `web-agent`, and its GitHub repository was published this week under the company’s account. The README says the package is an “open-source web data agent” built for structured web research. (github.com) The framework is designed around three core web tasks: search to find pages, scrape to pull clean content, and interact to operate sites through browser automation. Firecrawl’s docs describe the broader platform the same way: “search, scrape, and interact” as separate APIs for gathering web data. (docs.firecrawl.dev) Firecrawl said the agent is not tied to one model provider. Its launch post says users can “bring your own model,” choose a provider during setup, and deploy the project as a Next.js app, an Express server, or a plain library. (firecrawl.dev) That model-agnostic setup addresses a practical problem for teams building agents in 2026: browser automation, extraction logic, and retry handling often have to be rebuilt when they switch from one language model to another. Firecrawl positions its framework as the web layer underneath that model choice, not as a replacement for the model itself. (firecrawl.dev) Under the hood, Firecrawl says the agent uses Deep Agents from LangChain as its harness, which handles the plan-and-act loop, parallel sub-agents, and skill loading. Firecrawl’s own layer adds web tools, structured output, and streaming on top of that runtime. (github.com) The company is also tying the framework into the Model Context Protocol, or MCP, a standard way to let models call outside tools. Firecrawl’s docs say its MCP server can connect web-scraping and browsing capabilities to MCP-compatible clients, and its Google Agent Development Kit guide shows that same pattern for external agent frameworks. (github.com) (docs.firecrawl.dev) Firecrawl already sells a hosted web-data service alongside its open-source code. Its main repository describes the product as an API to “search, scrape, and interact with the web at scale,” while the company site says the service is available both as open source and as a hosted offering. (github.com) (firecrawl.dev) The immediate test is whether developers adopt the framework as a base layer instead of wiring together separate search APIs, scrapers, and browser bots on their own. Firecrawl’s pitch is that the agent should take a web task described in plain language and handle the rest. (firecrawl.dev)
