Firecrawl

Firecrawl's Alexandria is a unified data layer for AI agents, giving them one interface to discover sources, understand what each one holds, and retrieve information from the live web, third-party providers, custom connectors, and Firecrawl's own indexes. Those indexes include a Research Index covering tens of millions of scientific paper abstracts, a Developer Index spanning primary sources across documentation, READMEs, issues and merged pull requests, and a Government Index covering laws, regulations, and ordinances. The pitch to developers is that an agent researching a company can move from its website to filings, code, scientific literature, and enrichment data without wiring up a new API each time the question changes, reached through Firecrawl's existing API, MCP server, or CLI. The launch is paired with a $75 million Series B led by Smash Capital, with participation from Altos Ventures, Nexus Venture Partners, Y Combinator, Freestyle, and Offline Ventures. Firecrawl is best known as web data infrastructure used by more than 1.5 million developers, and Alexandria extends that footprint from crawling into a broader knowledge marketplace. A notable part of the plan is paying data providers directly, building on existing arrangements like Wikimedia Enterprise, through which millions of Wikipedia requests already flow each month , with a self-service option planned so individual creators and organizations can earn when agents use their data. The company was founded by Brazilian entrepreneur Nicolas Silberstein Camara alongside Caleb Peffer and Eric Ciarla , the same team behind Mendable, an earlier document-chat product. This launch matters right now because agent builders are hitting the ceiling of general web search and stitching together dozens of specialized data sources by hand, and Alexandria is Firecrawl's bet that the retrieval layer, not the model, is where the next round of quality gains will come from. Firecrawl reports internal evaluations showing agents using Alexandria scored 21% higher on answer quality than those relying on built-in web tools across 845 blind-judged tasks, a claim founders and operators evaluating agent stacks will want to test against their own workloads.