Firecrawl

An agent gets the web as tools: search, scrape, crawl, map, parse, monitor and Firecrawl's research agent, through its API, SDKs, CLI and an MCP server, and it is listed with LangChain, LlamaIndex, n8n, Make and Zapier. It connects to websites, not to the customer's other systems: there are no native connectors that act in a CRM, ticketing or messaging tool, and no per user credentials to scope, rotate or revoke for them. Acting inside a web page happens through the hosted browser rather than through connectors.

Sourcefirecrawl.dev/llms.txtread 2026-09-22

No workflow runtime for the customer is documented. The /agent endpoint searches, navigates and gathers data on its own from a prompt, optional URLs and a schema, but that pipeline is Firecrawl's own ready-made agent; the customer cannot define, branch or sequence its steps. Crawl, batch scrape and monitors are jobs Firecrawl runs to a fixed pattern. Nothing lets the customer sequence, branch, retry or route work while mixing deterministic nodes with agent steps; the customer's own logic lives in its framework or in tools such as n8n.

Sourcedocs.firecrawl.dev/features/agentread 2026-09-22

The public web is what Firecrawl searches, scrapes and crawls, and its Research and Developer indexes cover public papers and developer artifacts; none of it is a structure Firecrawl maintains over the customer's own documents. Parse turns a customer's PDFs and Office files into markdown and JSON for the customer to use, which is extraction, not an index that persists and stays queryable. Nothing documents Firecrawl indexing, refreshing or permissioning customer content. Clean output feeding the customer's own RAG layer is that layer's grounding, not Firecrawl's.

Sourcefirecrawl.dev/llms.txtread 2026-09-22

Threat Protection lets an organization set a policy, once, that every URL a request would fetch is checked against before Firecrawl fetches it: a scrape target, a search result, a link found during a crawl, or an agent's starting URL.

The policy takes a custom blocklist and allowlist of domains, blocked top-level domains, a risk score threshold and a fail-closed setting, can be driven by the organization's own Zscaler categories, and can be locked so no request weakens it. That is a structured guardrail and policy constraint, enforced before the action. No approval step, checkpoint or escalation to a person is documented. It is an enterprise feature enabled per organization.

Sourcedocs.firecrawl.dev/features/threat-protectionread 2026-09-22

SOC 2 Type II, independently audited, is listed in Firecrawl's enterprise documentation, and its security page carries the SOC 2 Type 2 mark. The customer-facing controls are documented beside it: SAML and OIDC single sign-on through Okta, Entra ID or Google Workspace, SCIM provisioning and deprovisioning, team roles, API keys restricted to IP ranges or locked to specific endpoints and formats, Zero Data Retention for scrape and search, and PII redaction, with DPAs and custom contracts. The enterprise controls are provisioned per organization.

Sourcedocs.firecrawl.dev/enterpriseread 2026-09-22

What the jobs and agent did is recorded one step at a time: SIEM Audit Logging streams one structured event for every URL Firecrawl fetches on the customer's behalf, whether from a scrape, crawl, batch, search, extract, parse or agent run, with the target, status, outcome (including fetches blocked by policy), the API key and workflow that caused it, a request identifier that groups a whole crawl or agent run, and the customer's correlation IDs, delivered to Microsoft Sentinel.

The dashboard's Activity Logs show request history with status, duration and credits. That gives an audit trail separate from the runtime, exported to a SIEM, covering each tool call. Page content is not logged, and retention by plan is not documented. SIEM logging is an enterprise feature enabled per organization.

Sourcedocs.firecrawl.dev/features/siemread 2026-09-22

No memory the agent reads to decide is documented. The closest thing is the saved login: a scrape can save its browser session to a named profile that later scrapes and Interact sessions load, and Browser Sandbox sessions can persist.

Cookies, local storage and an authenticated profile are loaded by the browser when a session starts, so the product applies them as settings rather than the agent reading them as context. No context persisted across a run, conversation or workflow and no longer term memory layer is documented, and nothing lets memories be reviewed, edited, deleted or scoped. Change tracking's comparison with the previous scrape is likewise applied as a rule.

Sourcedocs.firecrawl.dev/features/interactread 2026-09-22

Customers can use the managed cloud service or run Firecrawl on their own infrastructure: the core is open source under the AGPL 3.0 license and ships a pinned Docker Compose self-hosting guide, and the docs set out what each mode includes. Self-hosting exposes the core scraping APIs with the customer owning authentication, TLS, persistence, monitoring and upgrades and connecting its own OpenAI-compatible provider or Ollama for LLM formats, while Agent, Browser, Interact, the dashboard and enterprise controls are Cloud only. No region choice for the cloud service is documented.

Sourcedocs.firecrawl.dev/contributing/open-source-or-cloudread 2026-09-22

A single prebuilt agent ships: /agent searches, navigates and gathers structured data from a prompt and optional schema, also offered in the dashboard with no code; Firecrawl labels it a Research Preview in early access, and it is callable by any account. Firecrawl also publishes skills for Claude Code, Cursor and other coding agents covering Search, Scrape, Crawl and Interact. These are vendor made working assets for one job, web data gathering. One agent in preview and a skill set are not a browsable catalog of workflows, templates or role specific agents a buyer adopts.

Sourcedocs.firecrawl.dev/features/agentread 2026-09-22

Monitor runs recurring checks on a schedule the customer sets, against named URLs, a scheduled crawl of a whole site, or a web search, with an optional plain-language goal; each check labels every page same, new, changed, removed or error, and notifications go to a webhook per page, a webhook per completed check, an email summary or a Slack channel. Work starts on a schedule with no person initiating it. Crawl, batch and agent jobs also post lifecycle events to the customer's webhook.

Sourcefirecrawl.dev/llms.txtread 2026-09-22

Web data goes to whatever model the customer's agent runs, through LangChain, LlamaIndex, Mastra, the Vercel AI SDK and any MCP client, and a self-hosted deployment connects its own OpenAI-compatible provider or Ollama for LLM-backed formats. On the cloud service its own extraction and /agent models are Firecrawl's, with a choice among its agent models. Model choice sits with the customer for its own agent, and for extraction when self hosted.

Sourcedocs.firecrawl.dev/contributing/open-source-or-cloudread 2026-09-22

A documented REST API makes Firecrawl callable from outside, with published OpenAPI specifications for v2 and for webhooks, official SDKs for Python, Node, Go, Java, Ruby, Rust, .NET, PHP and Elixir, a CLI, and an MCP server that works keyless, with OAuth sign-in or with an API key, plus skills for coding agents.

Sourcedocs.firecrawl.dev/llms.txtread 2026-09-22

Firecrawl's Playground lets a developer try a request and inspect the raw response, and the Activity Logs' Debug issue button runs Firecrawl's support agent on a failed job to suggest a fix; both test and debug Firecrawl calls, not the customer's agent. No fixtures, test datasets, scoring or quality gates for the customer's agent workflows are documented.

Sourcefirecrawl.dev/llms.txtread 2026-09-22

Firecrawl hosts real browsers an agent controls: Browser Sandbox gives a secure, isolated session where agents fill out forms, click buttons and authenticate, with Playwright and agent-browser preinstalled, a CDP URL and a live view, and Interact continues on any scraped page by prompt or by Playwright code for multi-step flows such as signing in and clicking through pagination. That is documented control of a real interface, managed and sandboxed remotely. Browser Sandbox and Interact are Cloud only.

Sourcedocs.firecrawl.dev/features/browserread 2026-09-22