AI Search10 min read

What is MCP? How AI assistants connect to your tools (with Insight Engine examples)

MCP lets AI assistants like Claude and ChatGPT use outside tools and data. Here's what it is in plain words, how it works, and five real ways to use it with Insight Engine to track and improve your AI visibility.

By Insight Engine Team
Illustrated Insight Engine blog share image with a sample AI visibility scorecard with lime bars for ChatGPT, Claude, Gemini, Grok, and Perplexity, labeled SAMPLE, beside the title “What is MCP? How AI assistants connect to your tools (with Insight Engine examples).”

Most people use an AI assistant like a well-read stranger: it answers from what it already knows or can find on the web, but it can't see your accounts, your numbers or your tools.

MCP changes that. It's the reason Claude can now read your Google Drive, ChatGPT can look up a ticket in your project tracker, and, from today, either of them can pull your AI Visibility Score from Insight Engine and tell you what to fix this week.

Here's what MCP is, how it works, why it matters for your AI visibility, and five real conversations using Insight Engine's MCP server.

What MCP is, in plain words

MCP stands for Model Context Protocol. It's an open standard, first published by Anthropic in late 2024 and now supported by OpenAI, Google, Microsoft and most AI coding tools, that lets an AI assistant use outside tools and data in a consistent way.

The simplest analogy is a power socket. Before standard sockets, every appliance needed its own wiring. MCP is the standard socket: a service like Insight Engine builds one MCP server, and any assistant that supports MCP can plug into it. Nobody builds a separate integration for each assistant.

In practice, you ask your assistant a question in normal language and it fetches the real answer from a tool you've connected, rather than guessing.

How it works

There are three parts.

  • The client is the AI assistant you're talking to: Claude.ai, ChatGPT, Claude Code, Cursor or VS Code.
  • The server is the service being connected. Insight Engine's server lives at https://insightengine.studio/api/mcp.
  • The tools are the specific things the server offers, each with a name and a short description. Insight Engine offers ten, such as get_score and run_check.

When you connect a server, the assistant reads the list of tools and their descriptions. From then on, the assistant decides when to call a tool, based on what you ask. If you say "how are we doing in ChatGPT?", it works out that get_ai_answers is the right tool, calls it, gets data back, and writes you an answer in plain language. Most assistants show you which tool they called, and many ask before calling one that does something rather than just reading.

You approve access first. Connecting is a sign-in, not a password handover. With Claude.ai and ChatGPT you sign in to Insight Engine with Google and see a consent screen listing exactly what the assistant will be able to do. With developer tools you can use an API key instead, which you create and can revoke.

Why it matters if you track your AI visibility

Insight Engine checks how ChatGPT, Claude, Gemini, Grok and Perplexity answer your customers' questions, scores your visibility, and tells you what to fix. That's useful on a dashboard. It's often more useful where you already work.

Many people now spend part of their day inside an AI assistant, drafting emails, planning content or writing code. With MCP, your AI visibility data comes to that conversation. You can ask follow-up questions in your own words and turn a finding into a draft in the same place. An agency can pull every client's numbers in one request instead of clicking through fifteen workspaces.

What Insight Engine's MCP server can do

There are ten tools. Reading is free. Two tools use your plan's monthly allowance, and the assistant is told to say so before calling them.

ToolWhat it doesCost
list_sitesYour sites, each with its latest AI Visibility Score and last check dateFree
get_scoreThe score and its parts, the score per assistant, recent history, site health, and your Search score if Google Search Console is connectedFree
get_ai_answersWhat each assistant said in the latest check: excerpts, claims, the questions asked, which competitors were named instead, share of voiceFree
get_fixesFix checks (pass, warn or fail), next moves ranked by effect and effort, and fixes already generatedFree
get_usageChecks and generated fixes used this month against your planFree
run_checkRuns a fresh check across the assistants, taking one to two minutesOne of your monthly checks (max 6 a day)
get_runThe status of a check, and the new scores once it's doneFree
get_latest_reportYour newest monthly report, as data and markdown, with a PDF link that works for 24 hoursFree
generate_reportBuilds the monthly report now from the last 30 days of checksNo AI cost (30 a day)
generate_fixWrites three title and meta description rewrites, or a content brief for a pageOne of your monthly generated fixes

None of the tools edit your website. They read, check, report and draft. You decide what to publish.

Five conversations, start to finish

The numbers and business names below are examples, not real results.

1. The Monday check-in

You: How's our AI visibility this week, and what's the one thing I should fix?

Tools called: list_sites, then get_score and get_fixes for your site.

What comes back: Your AI Visibility Score is 46, up from 41 two weeks ago. You're named most often in Perplexity and least in Gemini, where you weren't named in any of the questions asked. Site health is fine. The top-ranked next move is a failing check: your homepage doesn't say plainly what you do and where, so assistants struggle to match you to category questions.

Cost: nothing; every tool here only reads.

2. "Run a fresh check and send me the report PDF"

You: We changed the homepage on Friday. Run a fresh check, then give me the report PDF.

Tools called: run_check, then get_run a few times while it runs, then generate_report.

What happens: Before running the check, the assistant tells you it uses one of this month's checks (it can call get_usage to show how many you have left). The check takes a minute or two. get_run returns the new score and the score per assistant once it has finished. generate_report then builds the monthly report from the last 30 days, including the new check, and returns it as markdown plus a PDF link anyone can open for 24 hours.

"Send" here means the assistant gives you the link; Insight Engine doesn't email it. If your assistant also has an email connector, it can draft the email.

Cost: one monthly check. The report itself has no AI cost.

3. What ChatGPT says about you, and who it names instead

You: What does ChatGPT say about us compared with Harbour Bookkeeping?

Tool called: get_ai_answers with the engine set to ChatGPT.

What comes back: Excerpts and claims from ChatGPT's answers in your latest check, the customer questions asked, who was named instead of you on each, and share of voice. The assistant summarises: ChatGPT named Harbour Bookkeeping in four of the questions where it didn't name you, described you accurately when it did mention you, and never named you first.

Two caveats. These are stored answers from your latest check, using your tracked questions; the tool doesn't ask ChatGPT anything new on the spot. And a competitor that never came up in those answers won't appear.

Cost: nothing.

4. Rewrite the pricing page title and meta

You: Our pricing page is at /pricing. The title is "Pricing | Northside Books" and there's no meta description. Write better options for "bookkeeping prices Bristol".

Tool called: generate_fix with the kind title_meta, passing the search, the page URL and the current title.

What comes back: Three title tag and meta description rewrites for that search, saved to your account alongside fixes made in the dashboard.

The tool doesn't read the page itself, so give it the current title and meta. It uses one of this month's generated fixes even if generation fails, so the assistant should confirm first.

Cost: one monthly generated fix.

5. An agency pulling every client's report

You: For each client site, get the latest monthly report and give me a table: site, score, change since last month, top next move.

Tools called: list_sites, then get_latest_report for each site. If a site has no report yet, the assistant can call generate_report for it.

What comes back: One table across all your clients, built from each report's scores, trends and next three moves, plus a 24-hour PDF link per client. In Claude Code you can save it straight to a file.

If you want a client's own app or dashboard to show their numbers, create an API key restricted to that client's site. The key can't see any other site on your account.

Cost: nothing, unless you generate reports (no AI cost, 30 a day).

Setting it up

The Connect page in the dashboard has the server URL and ready-to-copy snippets. The short version:

Claude.ai

Open Settings, then Connectors, and choose Add custom connector. Name it Insight Engine, paste https://insightengine.studio/api/mcp, choose Connect, sign in with Google and approve.

ChatGPT

Open Settings, then Apps and Connectors, and create a connector (developer mode, under Advanced). Paste the server URL, choose OAuth, sign in and approve.

Claude Code

Run this in your terminal, then type /mcp in Claude Code to sign in:

claude mcp add --transport http insight-engine https://insightengine.studio/api/mcp

For scripts or shared machines, create an API key on the Connect page and add it as a header instead:

claude mcp add --transport http insight-engine https://insightengine.studio/api/mcp --header "Authorization: Bearer <your key>"

Cursor

Add this to ~/.cursor/mcp.json, with your API key:

{
  "mcpServers": {
    "insight-engine": {
      "url": "https://insightengine.studio/api/mcp",
      "headers": {
        "Authorization": "Bearer <your key>"
      }
    }
  }
}

VS Code

Add this to .vscode/mcp.json:

{
  "servers": {
    "insight-engine": {
      "type": "http",
      "url": "https://insightengine.studio/api/mcp",
      "headers": {
        "Authorization": "Bearer <your key>"
      }
    }
  }
}

Your own apps

For an internal dashboard or a client portal, use the REST API at /api/v1 with an API key. It runs the same code as the MCP tools, with the same costs and limits. The API documentation has every endpoint.

What the assistant can and can't do

  • It only gets what you approve. The consent screen lists each permission in plain words, and the ones that use your plan allowance are marked with what they cost. API keys get read permissions by default; running checks and generating fixes are separate boxes you tick.
  • You can disconnect any time. Connected apps and API keys are listed on the Connect page, with an activity log of recent calls. Disconnecting or revoking takes effect straight away.
  • Keys can be limited to some sites. A site-restricted key can't see anything else on your account. A connection made through Claude.ai or ChatGPT acts for your whole account, so use a restricted key when you're giving access to someone else.
  • AI answers are passed along as quoted content. What ChatGPT or Gemini said about you is third-party text. The server labels it that way, so your assistant treats it as information to report, not as instructions to follow.
  • The paid plan is checked on every call. There's no cached pass. If a subscription lapses, calls stop until it's active again.
  • It can't spend more than the dashboard can. Checks and fixes come out of the same monthly numbers.

Try it

If you're on a paid plan, open the Connect page, copy the server URL and ask your assistant the Monday check-in question above. If you're not, pricing shows what each plan includes; the MCP server comes with all of them. And for the full list of what an agent can do, see the MCP server page.

Want to see this for your own site? Run a free SEO scan: https://insightengine.studio/scan

MCPModel Context ProtocolAI agentsAI visibilityClaudeChatGPT

Frequently asked questions

See how AI and Google see your site

Get your AI Visibility Score in about a minute: which assistants name you, who they name instead, and the fixes that raise it. Free plan, no card needed.

Paid plans add all five assistants every week. See pricing

built by 920four