Firecrawl Combines Skills with Search Sources, Letting Agents Choose Skill Summaries from Results

On November 4, Firecrawl merged skills search into its main branch, adding a skills source to POST /v2/search. Callers can use sources: ["web", "skills"] to search for web pages and skill summaries at the same time; summaries are returned in data.skills. This change gives agents a way to discover skills. Full skills must still be fetched separately, as no new skill-reading endpoint was added. Merge record and design notes

Each summary card includes an identifier, name, description, and source, and may also include a provider and similarity score. Skill searches return at most three cards, regardless of whether the general search limit is higher. Skill queries run in parallel with other sources and wait for up to three seconds. Timeouts, unsuccessful status codes, and unexpected responses are all converted to an empty array so that other search results can still be returned. The engineering trade-off is that applications cannot infer “no relevant skills” from an empty array alone; they must also account for possible service failures. Implementation constraints

This design extends Firecrawl’s existing tool-discovery workflow. The official Alexandria documentation describes a process of “search for tools, inspect inputs and pricing, then execute”; search results can include a provider, capabilities, and a description. Architecturally, adding skill summaries could let an agent review a small set of candidates before deciding whether to load their full contents. That is an inference about the integration pattern; no public performance evaluation has shown that it improves task success rates. Official Alexandria documentation

There are two further deployment constraints: skills search requires an API key and does not support zero data retention mode. The query itself is free, but that does not mean subsequent tool execution is free. At the time of review, the public search guide still did not list skills, so a merge to the main branch does not by itself confirm support across all cloud accounts or released SDKs. Teams integrating the feature should check the actual version and response fields, handle empty results, and test candidate quality on their own Chinese-language tasks. Search guide