firecrawl/web-agent

What it solves

Building a research-grade autonomous agent that can search, scrape, and interact with the web is complex. This project provides an open-source foundation so you can build your own agent instead of starting from zero. It bundles the core orchestration, tools, and templates.

How it works

The agent runs in a plan-act-observe loop: it plans, acts, observes, and repeats until the task is done. The loop is provided by LangChain's Deep Agents harness, which also handles sub-agent spawning and skills loading. Firecrawl's tools (Search, Scrape, Interact, bash) are wired into that loop. You can add reusable skills (SKILL.md playbooks) and spawn parallel subagents for independent tasks. Results can be returned as structured JSON via formatOutput, and the agent supports streaming.

Who it’s for

Developers who want to build their own web research agents or integrate autonomous web research into their applications. It's useful if you want to fork a working foundation, swap models, add skills, and deploy on your own infrastructure. It offers templates for a chat UI (Next.js), an API server (Express), or a bare library.

Highlights

• Ready-to-use Next.js and Express templates with chat UI, streaming, Skills, Subagents, and structured output
• Agent Core orchestration built on LangChain's Deep Agents, providing the agent loop, sub-agents, skills, and context management
• Firecrawl AI SDK integration: Search, Scrape, and Interact as Vercel AI SDK tools
• Structured output via formatOutput and data processing via bash tools (just-bash)
• Open source under MIT license