Web

Web is a CoreSpeed-hosted capability for grounding: it hands an agent current source material and leaves the reading to the agent. It needs no third-party account and runs behind the same /mcp seam as every other tool, so identity, organization holds, API-key caps, and activity recording apply unchanged.

Two tools

Tool	Use it for
`web__search`	Ranked results with source URLs for a query, each with a short text snippet.
`web__scrape`	The readable text of one `https` page, freshly crawled.

Search returns sources, not an answer

web__search does not synthesize. It returns ranked results — title, URL, published date, author, and by default a text snippet capped at 2,000 characters per result — and expects the agent to read them.

Parameter	Effect
`query`	What to search for. Required.
`num_results`	1–25, default 8.
`category`	Bias results toward one content type: `company`, `research paper`, `news`, `pdf`, `github`, `tweet`, `personal site`, `linkedin profile`, or `financial report`.
`include_text`	Inline a capped snippet per result. Default `true`.
`include_highlights`	Add query-relevant highlight sentences per result. Default `false`.
`livecrawl`	Force a fresh crawl of the results instead of the index cache — slower, fresher. Default `false`.

structuredContent

{
  "query": "MCP authorization spec RFC 9728",
  "resultCount": 2,
  "results": [
    {
      "title": "Authorization — Model Context Protocol",
      "url": "https://modelcontextprotocol.io/specification/…/authorization",
      "publishedDate": "2025-06-18",
      "author": null,
      "text": "…",
      "textTruncated": true
    }
  ]
}

textTruncated marks a snippet that hit the cap. To read a source in full, hand its URL to web__scrape.

Scrape reads one page

web__scrape fetches a single page and returns its clean text.

url must be https. IP literals, loopback, and internal hostnames are refused with invalid_input before any request leaves CoreSpeed.
max_characters defaults to 50,000 and is capped at 200,000. A clipped page sets textTruncated: true.
include_highlights adds query-agnostic highlight sentences from the page.

Every scrape is a fresh crawl. There is no multi-page crawl and no queue: for discovery across many pages, search first, then scrape the pages that matter.

Pricing and holds

Call	Charge
`web__search`	10 credits per request, whatever `num_results` is — the returned snippets are included.
`web__scrape`	5 credits per page.

A failed request is never charged. The live figures are repeated in the tool descriptions your client shows from tools/list, and every charge lands itemized in the organization ledger — see Billing & credits. An organization billing hold or suspension stops a web call before it runs, as an isError result carrying payment_required or org_suspended; API-key spend caps apply the same way.

Controls

Web appears as Web search & scrape in Dashboard → Capabilities. Turning it off hides both web__* tools from tools/list for that member; see Capability controls.

Two tools

Search returns sources, not an answer

Scrape reads one page

Pricing and holds

Controls

On this page