Prometheus
Prometheus is Firecrawl's tool for turning a plain-English data request into a reusable collector. It returns TypeScript code that uses the Firecrawl SDK, together with a sample of the data it produced. That gives a developer something to inspect and embed rather than only a one-time extracted answer.
Define the output before building
The documentation illustrates a request for Hacker News stories with titles, URLs, and points. A similarly precise request should name the target, required fields, and how much data to collect. Ambiguous requests leave the agent to make assumptions, which it reports in the returned explanation.
Prometheus says it experiments against the live site, writes a deterministic collector, and runs it before returning the result. The sample therefore describes the output from that build, not a guarantee that every later run will succeed or that the collected data is complete. Check important fields against the publisher's page and review the code before adding it to an application.
Keep permissions and source restrictions in scope. Generated collection code does not provide authorization to access private material or republish a site's content. A working collector can still request the wrong field or interpret a page differently from the person who wrote the prompt.
Builds, scripts, and deployments have different roles
A Build is the one-shot operation that returns code and sample data. A Script stores the collector with versioning and self-healing behavior. A Deployment controls when a script runs, using a schedule, an on-demand API request, or both.
When the target site changes, the documented repair process can append a corrected script version. Deployments that track the latest version adopt that update. Review those changes if downstream software depends on a stable schema; successful collection is different from preserving the exact interpretation your application expects.
Prometheus offers an HTTP API, CLI, MCP tools, and an installable agent skill around the same API contract. Choose the interface that fits the existing system rather than maintaining separate collection logic for each client.
Budget for collection time and provider access
The documentation estimates roughly 30 to 180 seconds for builds and prompt-based script creation, and up to about two minutes for an on-demand deployment run. Those timings should guide timeout handling, not be treated as a service guarantee.
The generated code uses Firecrawl and requires the relevant account credentials when run independently. Check current usage costs and collection limits before scheduling frequent runs. Preserve useful sample output and test downstream parsing so a repaired collector does not silently change a production data feed.