Firecrawl has released MCP v3, connecting AI agents to more than 3,000 tools in its Alexandria catalogue. The practical change is a single connection that can search across providers, keep going when one comes up empty, and hand work to parallel subtasks.
Firecrawl Watch analysis
What happened
Firecrawl says MCP v3 is available today in Claude, ChatGPT and other systems that support MCP. The catalogue includes company and people data, research, government records, and Firecrawl’s Search, Scrape and Parse tools.
The company says provider discovery is about eight times faster than in v2, MCP search is about twice as fast, and rate limits are up to 60 times higher. It also says users get four times more MCP usage, with free usage refreshing daily on every plan. Agents can delegate subtasks that continue running in the background.
Our top picks
- One connection to 3,000+ tools
Agents can reach a broad catalogue without wiring up providers one by one. - A fallback when a source comes up empty
Firecrawl says its waterfall continues through providers after a “not found”, reducing avoidable dead ends. - More headroom for repeated searches
The company says rate limits are up to 60 times higher and MCP usage is four times greater. - Parallel subtasks
Agents can split work, run subtasks in the background and carry on while results arrive. - Faster tool discovery
Firecrawl says discovery is about eight times faster than in v2, so agents spend less time finding a tool.
Why it matters
An agent is only as useful as the information and actions it can reach. A shared catalogue, fallback between providers and parallel work address three familiar snags: fragmented connections, missing results and slow runs. For teams doing research or batches of company lookups, that could make an agent less likely to stall halfway through the job.
The figures and performance comparisons are Firecrawl’s claims. Still, the features describe concrete changes to how an agent can find and use tools, rather than another promise that it will somehow become more capable by sheer force of branding.
Our read
This is a substantial update to the plumbing behind AI agents, and the fallback feature is the most interesting part: “not found” may mean “try the next source”, not “give up”. If you already use MCP-compatible agents, the daily free usage makes a small test easy. Check what providers and permissions your chosen setup actually exposes before handing it a larger task.
What to watch
- Whether the larger catalogue and higher limits hold up in everyday use.
- How often provider fallback finds useful records that the first source misses.
- Whether background subtasks make multi-step research more reliable.
Discussion spark: When an agent searches across thousands of tools, should convenience come first, or should users have to choose and inspect every data provider it can reach?
Sources and evidence
- Source update (12 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Firecrawl. Not affiliated with or endorsed by Firecrawl; this is not an official account.