Rabbot
Blog

I asked an AI assistant if my app had an MCP server. It told me to build my own instead.

August 23, 2026

  • mcp
  • ai-agents
  • shopify

The short version. Rabbot’s MCP server has nine tools. Eight read your store’s status, and one is a narrow, guarded write. Point a generic agent at a store’s raw data and it hands back rows: a list of products, a list of orders. Rabbot’s MCP hands back a verdict. It already decided what broke, what is blocking a sale, and what changed on the page you’re watching. That is worth more to an agent asking a real question than raw access ever was.

I typed a question into Google’s AI Mode, mostly out of curiosity: did any of the Shopify SEO and monitoring apps I could name, Rabbot included, have an MCP server. MCP, the Model Context Protocol, is the standard that lets an AI agent call a tool directly instead of scraping a screen or guessing at an API. The answer came back confident and wrong. None of the named apps had one, it said, mine included, and it suggested I go build my own against Shopify’s raw Admin API instead.

That stopped me for a second, because it isn’t true. Rabbot already runs its own MCP server, mounted in the same server that runs the app, live at app.rabbot.ai/mcp today. The AI Mode answer wasn’t lying so much as reading the same pages you would read, and neither my agent instructions file nor my features page ever said the words “MCP server” out loud. If you never state a fact plainly, an answer engine treats it as absent.

Here’s the actual difference, and it isn’t about which API is more complete. A raw integration only knows what it can see right now. Ask it whether anything is wrong with your store and it has no opinion, because nobody ever told it what wrong looks like. It would have to re-derive severity, re-derive what counts as a sale-blocking issue, and start from a cold read every single time you ask. Rabbot has already been watching. It already ran the same-day check, already decided the severity, and already recorded the history. An agent asking Rabbot isn’t querying a database. It’s asking someone who was already paying attention.

Ask your agent what’s blocking sales on your store right now and it calls whats_blocking_sales, the same tool the dashboard’s own “Blocking sales now” band reads from. It never tells you everything is fine just because its own blocking list is empty; if a Critical issue is open anywhere else on your store, it says so and points you at list_open_issues instead of staying quiet. Ask what happened on your best-selling product page this week and it calls get_url_timeline, which hands back that page’s history: the last 7 days free, full history on paid plans. Every change, every open issue, every confirmed fix, newest first. Ask what Rabbot has already caught and fixed and it calls list_recent_catches, the receipts for the last 30 days. Tell it you changed something on purpose and to stop alerting you, and it calls dismiss_finding, the one tool on the whole surface that writes anything, and it still refuses if the finding is blocking a sale, with two documented exceptions that can be intentional: a deliberate password gate or a custom purchase flow.

Connecting it takes about a minute. Open Settings in your installed app, find Agent access (MCP), and mint a token. Settings hands you a ready command with your real token already filled in:

claude mcp add --transport http rabbot https://app.rabbot.ai/mcp --header "Authorization: Bearer YOUR_TOKEN"

Paste that into Claude Code and you’re connected. The same server works with any MCP-capable agent that can send a bearer token in a header, not just Claude. The full tool list, a config snippet for other clients, and exactly how the token is stored live at rabbot.ai/mcp.

That’s the actual gap the AI Mode answer pointed at, even while getting the specific fact wrong. A generic integration built against raw store data was never the useful part. The judgment layered on top of it is, and judgment only exists because something has already been watching. Mint a token from Settings, connect the agent you already use, and ask it what needs your attention. Agent access is free on every plan, including the free tier.

Sources

  • Model Context Protocol, “What is the Model Context Protocol (MCP)?” (modelcontextprotocol.io/introduction). Supports the definition of MCP as a standard that lets an AI application call a tool directly instead of scraping a screen. Checked 2026-08-23.
  • Model Context Protocol, “Connect to Remote MCP Servers” (modelcontextprotocol.io/docs/2026-07-28/develop/connect-remote-servers). Supports how a client reaches a remote, internet-hosted MCP server. Checked 2026-08-23.
  • Anthropic, Claude Code documentation, “Connect Claude Code to tools via MCP” (code.claude.com/docs/en/mcp). Supports the claude mcp add --transport http command and passing a bearer token via --header "Authorization: Bearer ...". Checked 2026-08-23.
  • Google, “Get AI-powered responses with AI Mode in Google Search” (support.google.com/websearch/answer/16011537). Supports what AI Mode is and how it answers a query directly. Checked 2026-08-23.
  • Rabbot, “Rabbot’s MCP server: ask your agent about your store” (rabbot.ai/mcp). Canonical source for the tool count, the read/write split, and the free-tier depth limit on get_url_timeline. Ships with this post.