Your agent can reason. It can't see today's web.

#agents #ai #api #mcp

TL;DR: SpecStream is the live web data engine, one platform to build, search, extract, enrich, map and monitor data across the entire web. Its Search API is the most accurate web search API on the two standard benchmarks for search agents, at $5 per 1,000 searches. Sign-up comes with $10 of work, no card.

Try it: curl -G https://api.specstream.ai/v1/web-search -H "Authorization: Bearer $SPECSTREAM_API_KEY" --data-urlencode "q=Who is the CEO of Stripe?"

Your agent can reason. It can plan, call tools and write code.

But its knowledge stops the day its training did. Ask it about this morning and it has to go to the web, and the web it gets back is ten links, ten pages to read, and a guess about which one is right.

Models get better every month. The way they reach the live web hasn't kept up.

Duct tape as web access

If you're building agents today, web access probably looks something like this:

a search API that returns links
a scraper to fetch the pages
a model pass to read them
a second call when the first pages didn't have it

Four steps. It works. Kind of. Until the pages disagree, and the agent picks the first one it read.

SpecStream is the search your agent is missing

There's no index of our own. Every call reads the live web at the moment you ask. It asks several search engines at once, reads what the results print, opens the top pages when they disagree, and one model pass picks the answer.

You get the answer, the rows behind it, and the source URL and read time of every value, in a few seconds. When a question has many answers, you get them as rows. Add passages=true and the text of the top pages comes back in the same reply.

The benchmarks

On WideSearch, ByteDance Seed's benchmark for broad research tasks, SpecStream scores 60.9 with GPT-5.6 Luna. The best published score from another search API is Parallel's 57.6, and it needed the larger GPT-5.6 Sol. The small model on our search beats the big model on everyone else's.

On SimpleQA Verified, SpecStream scores 99 with GPT-5.6 Sol, against 97 for the best of the others.

Every score and the method: https://specstream.ai/benchmarks?src=devto

Pages, too

The Extract API reads 1 to 20 URLs in one call: web pages, PDF, Word and Excel. Each comes back as markdown or plain text with its own status, at $0.80 per 1,000 pages. A page with no text isn't charged.

Your agent finds it through MCP

The MCP server gives any MCP client two web tools, search_web and read_pages, plus Deep Search: ask in plain words and get every match on the live web in structured rows. For Cursor, in .cursor/mcp.json:

{
  "mcpServers": {
    "specstream": {
      "url": "https://api.specstream.ai/mcp",
      "headers": {
        "Authorization": "Bearer ${env:SPECSTREAM_API_KEY}"
      }
    }
  }
}

For Claude and ChatGPT, add a custom connector with https://api.specstream.ai/mcp and sign in. There are SDKs for Python and TypeScript, and the plain API works from anything that speaks HTTP.

Get started

Sign-up comes with $10 of work, no card. Then:

Search API: $5 per 1,000 searches
Extract API: $0.80 per 1,000 pages
MCP server: Claude, Cursor, ChatGPT, VS Code and more
Benchmarks: every score and the method

If you find a question it gets wrong, put it in the comments. I want to see it!

DEV Community