代理工具
Firecrawl MCP v3 Adds Provider Waterfall Queries and Background Subtasks, Expanding Agent Data Retrieval
Firecrawl released MCP v3 on November 12, adding provider waterfall queries and parallel background subtasks. The release integrates more than 3,000 tools, but feature parity between the hosted service and open-source server, as well as performance gains, still need to be verified separately.

On November 12, Firecrawl announced MCP v3, bringing together Alexandria’s tools for company, people, research, and government data with search, web scraping, and document parsing through a single connection. The update’s main technical change is that the service now handles some source switching and task dispatch, reducing the steps agents must take to orchestrate queries themselves. Official announcement
Provider waterfall queries follow a sequence of sources. For example, when enriching a person’s profile, if the first source has no record, the service continues to the next source. The agent needs to make just one call and receives the result. Inferred from this workflow, this could reduce the back-and-forth of having a model decide after each query whether to try another tool. However, the announcement does not explain source ordering, stopping conditions, or how conflicting data from multiple sources is handled, so it is not yet possible to determine whether data quality improves as well. Workflow details
Another new capability is background subtasks. According to Firecrawl, agents can split multiple queries into parallel tasks and continue with other steps while waiting for results. This could be useful for batch research and data enrichment, but engineering teams still need to validate task status reporting, failure recovery, and result aggregation. Firecrawl also claims that rate limits have increased by up to 60 times and usage by four times. These figures describe changes to service quotas; they do not directly indicate how much faster an end-to-end task will run. Changelog
Deployment requires clarifying which versions support which features. The public Firecrawl MCP Server README still lists 31 tools registered by the full configuration by default, while the hosted, keyless endpoint offers only three tools: scraping, search, and parsing. The release’s catalog of more than 3,000 tools should therefore not be taken to mean that every client will load an equal number of tool definitions at once, or that the local package already includes every v3 capability. Open-source server documentation
What remains to be seen is whether Firecrawl will publish complete v3 interface specifications, pricing and concurrency limits, and a feature comparison between the hosted service and self-hosted deployments. Before adopting it, teams can compare success rates, tail latency, and cost per valid result on a fixed query set to determine whether automatic source switching actually improves their existing agent workflows.