The hosted Firecrawl /agent is a research-grade autonomous AI agent powered by Spark 1 models optimized for structured web research. But every team wants something different: a different model, custom logic, their own infra.

Today we're releasing the open source firecrawl-agent stack so you can build exactly that.

This isn't a port of the hosted /agent — it's a lighter foundation designed to be forked and extended. One command scaffold a complete web agent project. From there, it's yours.

firecrawl-agent is an open framework for building powerful web agents. Bring any model. Anthropic, OpenAI, Google, or your own. It runs on Firecrawl's core primitives:

• /search for discovering pages,
• /scrape for extracting their content, and
• /interact for browser automation when the page requires it.

You get a plan-act agent loop with parallel sub-agents and structured output — and a working project you can run, modify, and deploy.

Two commands scaffold a full project:

# 1. Install the Firecrawl CLI and authenticate
npx -y firecrawl-cli@latest init -y --browser
 
# 2. Scaffold an agent project
firecrawl create agent -t next

You pick a template (Next.js, Express, or Library), pick a model provider, and get a project that runs immediately. The architecture is designed to be understood and changed, not treated as a black box.

Firecrawl handles the web layer. You own everything else.

Wiring an LLM loop, web tools, sub-agents, streaming output, and structured results from scratch takes time. firecrawl create agent scaffolds all of that — agent core, Firecrawl tool bindings, your choice of UI or API surface, and structured output.

You start from something that already works, then change what matters for your use case.

Three templates ship out of the box — pick the one that fits how you want to use it:

• Next.js — a full chat interface with streaming, for teams that want a UI to interact with the agent directly
• Express — a lightweight API server for integrating agent capabilities into existing tools, CRMs, or pipelines
• Library — a plain module for scripts and custom integrations, when you're embedding agent logic into an existing codebase

The firecrawl-agent stack doesn't lock you to a specific model provider. Swap in whatever LLM fits your budget, latency requirements, or data policy. Deploy to your own infra — there's no hosted dependency outside of the Firecrawl API itself for web access.

For teams running production pipelines — competitive intel monitors, lead enrichment, doc change detection — owning the full stack means you can tune, audit, and scale it on your terms.

The agent supports SKILL.md playbooks: markdown files that encode domain-specific procedures. Paginate through an e-commerce catalogue, reconcile data from multiple sources, apply a consistent extraction schema across a class of pages — anything your agent needs to do repeatedly, a Skill can encode.

Drop a Skills folder into your project and they're auto-discovered.

Three examples ship with the repo (e-commerce, deep research, structured extraction) as starting points. The value compounds as you write Skills for your own domain.

The plan-act loop can spin up parallel sub-agents for multi-target research jobs.

A competitive intel workflow covering 20 pricing pages, or a lead enrichment run across a list of companies, fans out across targets rather than working through them sequentially.

Results come back as structured JSON.

1. Install and scaffold:

npx -y firecrawl-cli@latest init -y --browser
 
firecrawl create agent -t next

2. Add your Firecrawl API key and model provider credentials
3. Pick a template, run the project, start customizing

For example workflows — competitive pricing monitors, lead enrichment pipelines, doc change detection — see the use case examples in the repo.

• A Firecrawl API key is required — the agent uses Firecrawl's hosted endpoints for scrape, search, and interact, even when self-hosted
• The plan-act loop and parallel sub-agents can consume significant tokens on complex jobs, so budget accordingly for model costs
• Browser automation via the interact tool is in beta and the API surface may change

firecrawl-agent is live and open source now.

Get started with firecrawl-agent · Read the docs