Grounded AI Answers MCP server

How an agent would call it

Two steps: search returns refs, then data_fetch opens one of them. The arguments below are the arguments that were sent.

{
  "tool": "search",
  "arguments": {
    "query": "What is the Model Context Protocol and who maintains it?",
    "sources": [
      "grounded-answers"
    ],
    "limit": 3,
    "runId": "<from run_begin>"
  }
}

{
  "tool": "data_fetch",
  "arguments": {
    "sourceId": "grounded-answers",
    "ref": {
      "id": "answer:What is the Model Context Protocol and who maintains it?"
    },
    "runId": "<from run_begin>"
  }
}

This runs against an account — send Authorization: Bearer <token> with the call. Get one at app.bedrocktap.com/dashboard/tokens; the free tier needs no card, and it reaches grounded-answers like every other source.

What it costs. 3 credits per question asked. Refs and fetches add nothing, so the 6 refs above cost 3 credits in total.

What came back

{
  "ref": {
    "id": "answer:What is the Model Context Protocol and who maintains it?",
    "title": "Grounded answer"
  },
  "text": "### What is the Model Context Protocol (MCP)?\n\nThe **Model Context Protocol (MCP)** is an open, standardized communication protocol designed to connect AI applications (large language models, AI assistants, and agents) with external data sources, business tools, and development environments.…",
  "fetchedAt": "2026-09-12T14:58:25.324Z"
}

One search returned 6 refs: the answer, then one per citation. The answer ref carried kind, answer, citationCount and searchQueries — the single query the model chose to run was "Model Context Protocol Anthropic maintenance", and it cited 5 sources. Each citation ref carried the publisher’s domain as its title: anthropic.com, modelcontextprotocol.io, medium.com and venturebeat.com. The answer itself ran to 2,233 characters.

Observed 2026-09-12. One call against a live source — ask again tomorrow and the rows move.

What a row contains

The first row is the answer: no URL, because a synthesised answer is not a page. Its meta carries the answer text, the search queries the model actually ran, and the citation count. Every row after it is one citation, carrying the source’s title. The citation URLs are search-service redirects rather than the publisher’s own address, so read the title to know who was cited. Grounding only fires when the model decides to search — a question it can answer from memory returns no citations, which is the model’s choice, not a fault.

A question it answers. What is the current state of this thing, and which pages say so?

Routing guidance

What an agent reads to decide whether grounded-answers is the right source, and what to reach for when it is not.

Answers

A first look at AI visibility on a budget: what a grounded assistant answers for a prompt, and which publishers it cites. Start here, then escalate to `ai-visibility` only if ChatGPT or Perplexity specifically matter.

Query shape

A natural-language PROMPT, not a keyword.

Price

Priced in credits, per row, per query or per call depending on the source. Every call is quoted before it runs — `cost_estimate` returns the exact figure and spends nothing. See /pricing.

Legal basis

Public API — the platform’s own documented, openly available interface.

Latency

Fast (~10s), against roughly 145 seconds for `ai-visibility`.

Prefer instead

`ai-visibility` when ChatGPT or Perplexity specifically matter. `google-serp` for the AI Overview shown on the results page — a different surface.

Cannot answer

What ChatGPT, Perplexity or the Google AI Overview say. This measures one engine and is not a proxy for the others.

Caveat

Cited links come back as redirect URLs, not publisher URLs. The publisher identity is carried separately — count share of voice on that, and resolve the redirect only when you need the canonical URL.

Caveat

Non-deterministic: one run is an observation, not a measurement.

Caveat

Bounded by a monthly allowance. Do not loop it over thousands of prompts without checking current pricing.

Related sources