Firecrawl Agent is an open source framework for building AI agents that autonomously research the web and turn what they find into structured data — as code you own. It was released on April 16, 2026 by Firecrawl (Mendable Inc.), the company known for its web scraping API. A single firecrawl create agent command scaffolds a fully working project from one of three templates: Next.js, Express, or Library. Under the hood it runs a plan-act-observe loop built on LangChain’s Deep Agents, calling Firecrawl’s Search, Scrape, and Interact primitives as its tools. You can swap in any model — Anthropic, OpenAI, Google, or your own — and deploy the result on your own infrastructure. It was formerly called Firecrawl Web Agent.
Key Features
- Scaffold a project with one command: Running something like
firecrawl create agent -t nextexpands a working project from a template. Next.js ships a streaming chat UI, Express gives you an API server, and Library is a plain module you can drop into scripts - Plan-act loop with web tools: Built on LangChain Deep Agents, it combines Search, Scrape, browser Interact, and bash execution so the agent can reach the information it needs without being handed URLs in advance
- Parallel sub-agents: Independent research targets are handed to parallel sub-agents with isolated sessions, so multiple sources can be worked at once
- SKILL.md playbooks: Domain-specific procedures written as Markdown files are auto-discovered and loaded on demand, letting you reuse the same research routine over and over
- Structured output: Specify a JSON schema and get results back in a fixed shape, ready to feed straight into a database or spreadsheet
- Choose your own model and runtime: The project’s stated goal is “bring any model” — Anthropic, OpenAI, Google, or your own all work. It runs against Firecrawl’s hosted endpoints or a self-hosted setup, and it is MIT licensed
Pricing
The framework itself is MIT licensed open source and free to use, modify, and redistribute. Actual costs come from two places: Firecrawl API credit consumption, and the API fees of whichever LLM provider you choose. Firecrawl’s own plans are below.
| Plan | Monthly price (billed yearly) | Credits/month | Concurrency |
|---|---|---|---|
| Free | $0 | 1,000 | 2 |
| Hobby | $16 | 5,000 | 5 |
| Standard | $83 | 100,000 | 25 |
| Growth | $333 | 500,000 | 50 |
| Scale | $599 | 1,000,000 | 100 |
| Enterprise | Contact sales | Custom | Custom |
These are monthly equivalents when billed yearly; month-to-month billing differs, so check the official site. Note also that Firecrawl’s hosted Agent API (the /agent endpoint) is a separate product from this framework — during its research preview it bills dynamically by credit and gives every user five free runs per day.
Pricing is current as of August 2026. Please check the official site for the latest pricing.
Pros & Cons
✅ Pros
- A single command gives you a working project, so the entire initial setup of agent scaffolding is skipped
- Being MIT licensed open source, you can read the code, understand it, and reshape it freely
- You pick the model, swapping between Anthropic, OpenAI, and Google to match your cost and quality requirements
- It deploys to your own infrastructure, which makes it viable in environments where handing data to an outside SaaS is not an option
- SKILL.md lets you accumulate procedures, turning research know-how into a team asset
⚠️ Cons
- This is a developer framework, not a no-code tool; TypeScript/Node.js knowledge is assumed
- “Free framework” still means Firecrawl credits plus LLM API fees, and costs get hard to predict as run volume grows
- Because it crawls the web autonomously, you have to design your own handling of target sites’ terms of service and robots.txt
- It is recently released, so case studies and third-party material are still scarce
- Its name is close to the hosted Agent API, which makes it easy to confuse which one is being discussed
Comparison with Similar Services
| Criteria | Firecrawl Agent | Firecrawl Agent API (hosted) | Browser Use | Stagehand |
|---|---|---|---|---|
| Form | Open source framework | Hosted API | Open source library | Open source framework |
| Main use | Building your own web research agent | Autonomous research from a prompt alone | Browser automation | Browser automation |
| Runtime | Your own infrastructure, anywhere | Firecrawl’s side | Your own infrastructure | Your own infrastructure / Browserbase |
| Model choice | Any (Anthropic/OpenAI/Google, etc.) | Firecrawl’s own models | Any | Any |
| Primary language | TypeScript / Node.js | Language agnostic (HTTP API) | Mostly Python | Mostly TypeScript |
Who Is It For
- Developers who want a self-built agent to handle recurring web research or competitive analysis
- Teams that cannot hand data to an outside SaaS and need an agent running in their own environment
- Anyone who wants to learn how LangChain Deep Agents are assembled by reading working code
- Developers moving existing scraping work from fixed-URL fetching to goal-directed autonomous research
- People who want a ready-made agent foundation while keeping control of their model choice
Summary
Firecrawl Agent packages the common architecture of a web research agent into templates and gives it away as open source. Its real value is that the fiddly parts — the plan-act loop, parallel sub-agents, reusable procedures via SKILL.md, structured output — are there from the start. That said, it is a developer-facing foundation, and costs accrue to both Firecrawl and your LLM provider in proportion to how much you run. A sensible approach is to run a template on Firecrawl’s free tier first, measure how many credits your expected research volume consumes, and decide on full adoption from there.