Firecrawl MCP server

Keep firecrawl-mcp and change one variable

firecrawl-mcp reads FIRECRAWL_API_URL, which its README names as the custom API endpoint. SpecStream answers Firecrawl's v2 scrape, search, map and crawl requests at https://api.specstream.ai/firecrawl in Firecrawl's own JSON. Set that address, put a SpecStream key in FIRECRAWL_API_KEY, and firecrawl_scrape, firecrawl_search, firecrawl_map and firecrawl_crawl run on SpecStream with your prompts unchanged. In claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-server-firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_URL": "https://api.specstream.ai/firecrawl",
        "FIRECRAWL_API_KEY": "YOUR_SPECSTREAM_KEY"
      }
    }
  }
}

firecrawl_agent, firecrawl_interact, firecrawl_parse, the monitor tools and the Alexandria tools call Firecrawl endpoints that SpecStream does not answer. For research across many sites, add SpecStream's own server beside it and use Deep Search.

10,000 pages for $8

On Firecrawl's Hobby plan, 10,000 pages take 10,000 credits, twice the 5,000 that $19 a month holds. On SpecStream they cost $8 from your balance, with no plan to buy. Firecrawl charges 1 credit for a page that answers 403 or 404 and 5 credits for a page in JSON format. On SpecStream a page that gives no text costs nothing, and every page that does costs $0.0008.

The most accurate web search API in the world, behind both servers

firecrawl_search and search_web both run SpecStream's Search API. It scores 97.9 F-score on all 1,000 official SimpleQA Verified questions (978 correct), and 64.0 item F1 and 43.4 row F1 on all 200 official WideSearch tasks. The Extract API behind firecrawl_scrape and read_pages scores 0.908 F1 on WCXB. The scores and the method are on the benchmarks page.

SpecStream's own server in one line

In Claude Code, run this and sign in when the browser opens. Nothing runs on your machine:

claude mcp add --transport http specstream https://api.specstream.ai/mcp

Your agent gets search_web for search, read_pages for up to 20 pages or documents a call, and Deep Search for every match on the live web as structured rows, with the source URL and read time on every value. Cursor, VS Code, ChatGPT and the other clients take the same address; the MCP client guides show where each one keeps it.

Tavily MCP

Tavily's hosted server at https://mcp.tavily.com/mcp/ offers search, extract, crawl, map and research. Basic extract costs 1 credit for every 5 pages, $1.60 per 1,000 pages at $0.008 a credit, and a basic search costs 1 credit, $8 per 1,000. SpecStream reads pages at $0.80 per 1,000 and searches at $1 per 1,000.

Exa MCP

Exa's server at https://mcp.exa.ai/mcp gives web_search_exa and web_fetch_exa by default. Search starts at $4 per 1,000 requests and page contents cost $1 per 1,000 pages. SpecStream's read_pages costs $0.80 per 1,000 pages, and its search API scores 97.9 F-score on SimpleQA Verified.

Apify MCP

Apify's server at https://mcp.apify.com lets an agent find and run Actors from Apify Store. The free plan gives $5 to spend in Apify Store, and compute costs $0.2 per compute unit, so the price of a page depends on the Actor you run. SpecStream prices every page the same, $0.0008.