MCP v3 is out from Firecrawl. It’s the company’s latest Model Context Protocol server for AI agents. A 12 November blog post and a matching changelog entry say you can use it now in Claude, ChatGPT and any other app that speaks MCP.
According to Firecrawl, one connection now gives an agent more than 3,000 tools, automatic provider waterfalls, up to 60 times higher rate limits, four times more usage through the MCP, and subtasks that can run in parallel. The company also published a landing page for the demo.
Introducing Firecrawl MCP v3. Give your agents access to all of the knowledge they need. → 3000+ tools with automatic provider waterfalls → 60x higher rate limits → 4x more usage → Spawn subtasks for parallel work. Available now for every harness.
November 12, 2026
What the 3,000 tools are
Firecrawl says the MCP hooks an agent into the full Alexandria catalogue of structured-data providers, covering company data, people data, research and government records, with Search, Scrape and Parse right there too. You tell the agent what you need, and the MCP is meant to pick the tool. The company puts provider discovery at about eight times faster than in v2.
The 3,000-plus tools sit in that catalogue. The model doesn’t get all 3,000 definitions stuffed into its context. What the hosted server actually lists is a smaller MCP toolset: search, scrape and parse, and browsers such as firecrawl_find_tools. In the open-source 3.40.0 changelog, dated 12 November, Firecrawl also added alexandria_library_search and alexandria_library_read, which treat the catalogue like a read-only library of provider contracts. The same plugin already carries Firecrawl’s job search tools, which run on Alexandria job and people data.
I opened that library with the free catalogue tools. They don’t run providers or spend credits. I got 36 categories, 574 data providers and 3,294 capability files. If each capability is a tool, that backs the “3,000+” claim. Some providers are company and people-data firms; others are official sources, including Wikipedia and the Federal Reserve’s FRED database.
When the first source comes up empty
Firecrawl calls it a waterfall when the system keeps going after the first “not found.” The blog uses people enrichment as the example: one call tries a source, then the next, until it finds the record. You hook up once, Firecrawl says, instead of adding providers by hand.
Version 3.40.0’s changelog is more specific for people and company search and enrichment. It tries providers one at a time in listed-price order, and for enrichment it also looks at past success. The first provider to find the field wins. You pay the listed price on a hit; a miss costs nothing. Leave max_credits off and an enrichment call has no cap. Search rows come back as name, title, company, location, LinkedIn URL and source, with no emails or phone numbers. If you want contact details, enrichment can ask for them, including work email. Mobile is opt-in. The result tells you which provider found each field.
A second source can fill a hole in the data. It still might be wrong. Enrichment can mark emails as verified, and verified_only will drop the rest.
| Feature | What Firecrawl says | What I could check |
|---|---|---|
| 3,000+ tools | The full Alexandria catalogue, plus Search, Scrape and Parse, through one MCP connection. | The free library browser listed 3,294 capabilities across 574 providers in 36 categories. |
| Provider waterfalls | If one source comes up empty, the next is tried. | For people and company search and enrichment, 3.40.0 tries providers one at a time by listed price and stops at the first hit. A miss is free; a hit is charged. |
| Rate limits and usage | Up to 60 times higher rate limits, four times more usage through the MCP, and free daily MCP usage on every plan. | The claims appear in the blog and launch video. The public rate-limit docs don’t give the baseline for those multipliers. |
| Parallel subtasks | An agent can spawn background subtasks and pick up the results as they land. | The blog describes this, and the launch video shows a Claude chat starting three Firecrawl agents, one each for New York, San Francisco and London. I didn’t find a matching how-to page in the public MCP docs. |
Limits, subtasks and how to connect
Firecrawl also says search through the MCP is about twice as fast, and that the extra headroom matters once an agent isn’t limited to one call at a time. The launch video ends on “free daily usage.” I read the rate-limit docs: they list plan caps, and daily request and credit caps per IP for keyless use, but they don’t say what the 60-times and four-times figures are measured against.
Without a key, the hosted MCP still only offers Search, Scrape and Parse. Sign in or add an API key and you can get the full tool surface, if your plan and team policy allow it.
Firecrawl says an agent can now hand off work and keep going, for instance by splitting research on ten prospects into ten jobs that run at once. I looked for a how-to on how those subtasks are started, billed or reported back, and didn’t find one in the public MCP docs, only the blog and the video.
The MCP docs cover setup. You can sign in through the browser at https://mcp.firecrawl.dev/v2/mcp-oauth, put an API key on https://mcp.firecrawl.dev/v2/mcp, or run the open-source server on your own machine. ChatGPT and Claude have a native Firecrawl connector. A user asked whether to hook the MCP up directly or use the official Codex plugin, which they said takes a long time to update. Co-founder and CEO Caleb Peffer said Firecrawl is working with OpenAI to get the plugin updating as fast as the MCP.
The new Firecrawl MCP is live 🔥
November 12, 2026
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
