Firecrawl launches Alexandria and raises $75M Series B

Firecrawl introduced Alexandria, a knowledge library that combines live web crawl with official data providers and Firecrawl indexes for AI agents, and raised a $75 million Series B led by Smash Capital.

SOFTWARE desk — same-day product and raise: the scrape layer is growing into a paid knowledge exchange so agents can cite sources humans still own.

What Alexandria is, in the post’s words. It brings official data providers, custom connectors, and Firecrawl’s own indexes together with the live web, so an AI agent has one way to find a source, see what it holds, and pull from it. An agent, here, is software that takes steps to answer a question, not only a chat reply. An index is a searchable collection Firecrawl already built, separate from whatever is on the open web at the moment of the question. A connector is a link to a source the customer brings. The post says the agent does that through the Firecrawl API already in use, or through Firecrawl’s MCP. An API is the service a developer’s software calls. MCP is the Model Context Protocol, the plug an agent uses to call a tool. The post says Alexandria starts today, on its own or next to web search and scraping. Scraping means copying a page off the public web. File that product description as the company’s. This desk did not call the API.

Why the company says the money buys knowledge. The post says a good chunk of the round will go to buying knowledge from people. Firecrawl already pays official data providers through agreements, including Wikimedia Enterprise. Wikimedia Enterprise is the paid service the Wikimedia Foundation runs so a company can use Wikipedia data under a contract, rather than only scraping the public site. The post says millions of requests for Wikipedia data flow through Firecrawl each month, and that Firecrawl pays for direct access, which it says supports Wikipedia and gives users a cleaner way to retrieve the pages. The funding, the company says, will extend that kind of paid access to more people and organizations, and will improve search and add sources agents can reach. A self-service system so individuals, content creators, and organizations can earn from what they know is something the company plans to open soon. That includes expertise that has never been posted online. Planned is not open. This desk did not see a contract or a creator payout.

The indexes the post names, and what those words mean. The Research Index includes tens of millions of scientific paper abstracts. An abstract is the short summary at the front of a paper, not the full paper. The Developer Index spans tens of millions of primary sources across documentation and READMEs, issues, and merged pull requests. Documentation is the written guide for a piece of software. A README is the explanation file that sits with a project. An issue is a bug report or a task. A pull request is a proposed code change that has been accepted. The Government Index covers laws, regulations, and ordinances. “Tens of millions” is the company’s phrase for each of the first two indexes. It is not a count this desk audited, and it is not a claim that every paper or every law is in the library.

The quality line is a company test, not a score this desk ran. Across the verticals Firecrawl tested, agents using Alexandria scored 21 percent higher on answer quality than agents using built-in web tools. A vertical, here, is one kind of question. The company says it used the same model and the same prompts across 845 tasks, with blind AI judging. A prompt is the instruction you give the model. Blind, here, means the judge was not told which system wrote the answer. The judge, on this page, is another AI, not a panel of people. The post does not name the model, the judge, or the verticals. File 21 percent and 845 as Firecrawl’s. Do not treat them as an outside ranking.

Where the crawl product came from, still the post. Before Firecrawl, the company built Mendable, an AI chat product for documentation. The post says Mendable worked, and that the hard part was getting clean, reliable information out of the web. Other teams building with AI were solving that same problem from scratch. Firecrawl is the tool they built for that job. Give it a web address and it crawls, renders, parses, and cleans the page. Crawl means visit pages and follow links. Render means load the page the way a browser would, including parts that appear only after the page runs. Parse means pull the text out of the layout. The post says they expected a few hundred people to need it. Over 1.5 million users build with Firecrawl now. That count is the company’s. This desk did not audit the user list.

What sits beside the launch, and is not a second result. The post walks through an example: an agent researching which startups already build a product, then looking up people who might buy it, without a new data connection for each question. That example is a story on the page. It is not a customer count, and it is not a measured win. The post says the next work is better access to the live web, deeper indexes and retrieval, and more first-party sources. First-party means the data comes from the organization that owns it. Retrieval means finding the passage that answers the question, not only the name of the source. The command that installs a command-line helper stays in Sources. It is not the product this card is filing.

Do not fill blanks the page left open. There is no valuation. There is no split of the $75 million among the named investors. There is no hour on the post. The self-service system for creators is a plan, not a system this desk can open. The 21 percent line is not a test this desk reran. “The library for superintelligence” is the company’s phrase for the ambition. Superintelligence, in that sentence, is the company’s word for AI that can draw on a very wide library. It is an ambition. It is not a result, and it is not a claim that such a system exists today.

Plain English for the rest of the card: Series B = a later private funding round. $75 million is the new money the post prints. It is not a price for the company. The page does not print a valuation. agent = software that takes steps, not only a chat reply. index = a searchable collection Firecrawl already built. connector = a link to a source the customer brings. API = the service a developer’s software calls. MCP = the plug an agent uses to call that service. scraping = copying a page off the public web. Wikimedia Enterprise = the paid Wikipedia data service. abstract = the short summary of a paper. README = the explanation file with a software project. issue = a bug report or a task. pull request = an accepted proposal to change code. vertical = one kind of question in the company’s test. prompt = the instruction given to the model. blind judging = the judge is not told which system wrote the answer. crawl = visit pages and follow links. render = load the page the way a browser would. parse = pull the text out of the layout. first-party = the organization that owns the data. This filing is the 22 Sep announcement. It is not an audit of the round.

PRIMARY here: Firecrawl’s 22 Sep 2026 post, “Introducing Alexandria and our $75M Series B,” by Caleb Peffer, at firecrawl.dev/blog/introducing-alexandria-series-b — Tier A PRIMARY, the company’s own record. The $75 million Series B, the Smash Capital lead, Altos Ventures, Nexus Venture Partners, Y Combinator, Freestyle, Offline Ventures, the Alexandria description, the API and MCP line, the Wikimedia Enterprise agreements, the millions of Wikipedia requests a month, the planned self-service system, the Research, Developer, and Government indexes, the 21 percent claim on 845 tasks with the same model and prompts and blind AI judging, the Mendable origin, the over-1.5-million-users line, and the “starts today” line are the post’s. NOT claimed: a valuation, a split of the $75 million, an hour stamp, that the creator system is open, that this desk reran the 845 tasks or audited the user count or the index sizes, that “superintelligence” is a result, a stock tip, or investment advice. Distinct from the already-filed openai-frontier-standards, aws-strands-harness, and meta-muse-0day.