Skip to content

/agent by Firecrawl Review

/agent is Firecrawl's web research API: describe the data you want in plain language, optionally add a JSON schema, and it searches, browses and returns structured JSON. Built for developers.

Free plan · from $19/mo · API

Stork's verdict on /agent by Firecrawl

/agent is the quickest way to turn a plain-language data request into JSON from pages you have not found yet, but set maxCredits and test on free runs first.

/agent by Firecrawl reviewed by Stork AI · stork.ai/en/agent-by-firecrawl

What /agent by Firecrawl does

/agent is an endpoint of the Firecrawl API (POST https://api.firecrawl.dev/v2/agent) that Firecrawl describes as deep research for data. You send a prompt (up to 10,000 characters) describing what you need, such as a list of companies with founders and websites, and the agent searches the web, navigates sites, handles JavaScript pages, pagination and infinite scroll, and returns JSON. You can pass a JSON schema (Pydantic in Python, Zod in Node) to get typed output, and an optional list of urls to focus the run; strictConstrainToURLs keeps it on those pages only. Firecrawl says /agent replaces its older /extract endpoint, which it is deprecating, and that for a single known URL its JSON mode on /scrape is cheaper.

Every run now uses Firecrawl's Spark 2 model (the older Spark 1 Mini and Pro names still work but route to Spark 2), with an optional effort setting of low, medium or high. Jobs run asynchronously: you poll by job ID or receive webhooks, results stay available through the API for 24 hours, and a thread ID lets you send follow-up prompts that continue a run. Other options include a proxy country for the whole run, SafeSearch filtering, Zero Data Retention (billed at 2x), and access to paid Alexandria data providers with an approval step. You can try it without code in the playground at firecrawl.dev/agent, or call it from the Python and Node SDKs, the CLI and Firecrawl's MCP server. Parallel Agents batches hundreds or thousands of /agent queries as a spreadsheet-style grid.

/agent by Firecrawl pricing

/agent has no plan of its own. Every account gets 5 free agent runs a day; beyond that, runs draw on the same Firecrawl credit balance as scraping and search. Firecrawl plans:

PlanMonthly priceBilled annuallyCredits per monthExtra credits
Free$0-1,000-
Hobby$19/month$16/month5,0001,000 per $5
Standard$99/month$83/month100,0002,000 per $5
Growth$399/month$333/month500,0002,500 per $5
Scale$749/month$599/month1,000,0005,000 per $5
EnterpriseCustomCustomCustomCustom

Agent pricing is dynamic: Firecrawl does not publish a fixed price per run. Its docs say "most agent runs consume a few hundred credits", with simple single-page tasks using less and multi-domain research using more, and that you can watch credit use live in the agent panel. As an illustration only, a 300-credit run costs about $1.14 on monthly Hobby ($19 / 5,000 x 300) or about $0.30 on monthly Standard ($99 / 100,000 x 300). In Parallel Agents, the Spark 2 Fast quick-lookup path costs 10 credits per cell, and cells that need deeper research are billed by tokens at your chosen effort.

Cost controls: maxCredits caps a run and defaults to 2,500 credits (the dashboard allows up to 2,500; higher values are set through the API and are always billed as paid). A run that hits its cap stops with stopReason: credit_limit_reached, returns no data (sometimes a partial result), and can be continued in the same thread. Zero Data Retention runs bill AI credits at 2x. Pay-as-you-go top-ups in $5 increments are available on paid plans; Scale rolls unused credits over for one month.

Prices checked on the vendor's pricing page on November 10, 2026. See Firecrawl pricing and the Agent pricing notes.

Key features

What Firecrawl's /agent page and docs list:

  • Prompt-only requests: no URLs needed, the agent searches and discovers sources
  • Structured JSON output against your own schema (Pydantic or Zod in the SDKs)
  • Optional urls list, with strictConstrainToURLs to stay on those pages
  • Spark 2 model with low, medium or high reasoning effort
  • Async jobs with polling, webhooks, execution traces and follow-up turns in a thread
  • Per-run maxCredits budget, with partial results when a run hits the cap
  • Proxy country per run, SafeSearch option and Zero Data Retention mode
  • Parallel Agents for batches of hundreds or thousands of queries

Who /agent by Firecrawl is best for

/agent suits developers and data teams who need structured data but do not know which pages hold it: lead lists, product and price catalogues across retailers, company research or one-off datasets. It is most useful when you would otherwise write a search step, a crawler and an extraction prompt yourself. If you already know the exact URL, Firecrawl itself recommends JSON mode on /scrape, which is cheaper and synchronous; see the main Firecrawl listing.

Because pricing is per run and varies, it is a weaker fit for teams that need a fixed cost per record. If you want a no-code scraper with ready-made robots, look at Browse AI or Apify. For a search API that returns web results for your own agent to process, compare Exa, Tavily or Parallel.

What users say

Independent hands-on reviews of /agent are scarce; most Firecrawl reviews cover the scrape and crawl APIs. A Firecrawl review on Thunderbit's blog, written by Thunderbit's CTO (Thunderbit sells a competing scraper), describes Agent as a fit for research "when URLs are unknown or navigation is required" and warns that dynamic pricing makes costs hard to predict, advising a maxCredits limit and testing with representative prompts. Firecrawl's own docs say the same: Agent "can be expensive", and they suggest splitting large jobs into runs of 3 to 5 URLs, since one run returns roughly 150 to 200 rows. Use the 5 free daily runs to measure credit use on your own prompts before relying on it.

Pros

  • Needs only a prompt; finds and navigates sources without a URL list
  • Returns JSON that matches your own schema
  • 5 free runs a day to measure credit use before paying
  • Hard per-run spending cap with maxCredits
  • Shares one credit balance with Firecrawl scrape, crawl and search

Cons

  • No fixed price per run; credit use varies by task
  • Firecrawl's docs say most runs use a few hundred credits
  • A run that hits its cap returns no final data, only a possible partial
  • Few independent hands-on reviews of /agent so far

Frequently asked questions

What is Firecrawl /agent?

It is a Firecrawl API endpoint that takes a natural-language prompt, searches and browses the web on its own, and returns structured JSON. URLs and a schema are optional. Firecrawl says it replaces the older /extract endpoint.

How much does Firecrawl /agent cost?

Every account gets 5 free agent runs a day. After that, runs use Firecrawl credits at a dynamic rate that depends on the task; Firecrawl's docs say most runs use a few hundred credits. Plans run from Free (1,000 credits) to Hobby at $19/month (5,000 credits) and Standard at $99/month (100,000 credits). Prices checked on November 10, 2026.

How do I limit what an agent run spends?

Set maxCredits on the request. It defaults to 2,500 credits. If a run hits the limit it stops with stopReason: credit_limit_reached, returns no final data (sometimes a partial result), and you can continue it in the same thread with a new budget.

What is the difference between /agent and /extract?

/extract needed URLs; /agent needs only a prompt and finds sources itself. Firecrawl lists /agent as faster, cheaper and more reliable than legacy /extract and says /extract is being deprecated.

What model does Firecrawl /agent use?

Every run uses Spark 2, Firecrawl's agent model, with an optional reasoning effort of low, medium or high. The older Spark 1 Mini and Spark 1 Pro names are still accepted but route to Spark 2.

Can I use Firecrawl /agent without code?

Yes. The playground at firecrawl.dev/agent and the agent panel in the Firecrawl dashboard run agents from a prompt. Developers can also call it through the REST API, Python and Node SDKs, the Firecrawl CLI and the MCP server.

How long does a Firecrawl agent run take?

Firecrawl says simple extractions finish quickly and complex multi-source tasks may take a few minutes. Runs are asynchronous, so you poll the job or receive webhook events, and results stay available through the API for 24 hours.

Compare Alternatives

/agent by Firecrawl

Free plan · from $19/mo · API

From the Stork blog

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking

Stork Quadrant

Becomes the API· 33/100

Replaceable as a UI, but kept alive as the API the agents call.

“Firecrawl is a well-executed API wrapper around capabilities that browser-equipped LLMs and open-source scrapers already cover. The brand is real — builders know the name — but brand alone at a 1 on the moat scale won't hold when every major LLM provider ships native browsing and structured extraction. The core value prop, clean web data for agents, is exactly what OpenAI, Anthropic, and Google are building into their own stacks.”

— Claude Sonnet 4.6, scored 2026-05-27

Defensibility · 7/100

  • Physical-world coupling
  • Regulatory moat
  • Network liquidity
  • Proprietary refreshing data
  • High-trust catastrophic workflows
  • Multi-party coordination
  • Brand / community / taste

An LLM alone could replace

  • Summarize or extract information from a webpage given a URL
  • Generate structured data from unstructured web content
  • Write a web scraping script in Python using BeautifulSoup or Playwright
  • Search the web and return relevant results with context

Agent-Readiness · 65/100

  • Verified MCP— Stork MCP listing: firecrawl-mcp (confirmed)
  • Listed on agent surfaces— anthropic_directory, cursor + Stork:firecrawl-mcp
  • Usage-based pricing
  • Headless agent auth— https://www.firecrawl.dev/docs (api-key auth)
  • Public OpenAPI
  • Active changelog
  • llms.txt— https://www.firecrawl.dev/llms.txt

Score history · +24 pts over 3 re-scores

How to defend

Go vertical: pick domains where scraping requires auth, anti-bot evasion at scale, or legal compliance (e.g., financial data, job boards, real estate) and own the reliability SLA there. Alternatively, become the coordination layer agents call — not just a scraper but the managed pipeline with caching, deduplication, and change detection that no LLM alone can replicate.

  • Add a usage-based or per-call tier; per-seat-only pricing dies when agents replace seats (+15).
  • Publish an OpenAPI spec at /openapi.json or /.well-known/openapi (+10).
  • Publish a public changelog and ship in the last 90 days — silence reads as abandonment (+10).