Kilo from your AI assistant (MCP)

MCP is to AI assistants what a REST API is to programs: a standard way in. Kilo’s MCP server lives at the same origin as the API - one hosted server, run by Kilo, nothing to install - and exposes a small set of tools an assistant can call on your behalf once you hand it an API key.

The practical effect: “make a QR code for the menu, and tell me next week how it did” becomes a complete workflow you can run from a chat window. The assistant creates the link, saves the QR styling, and reads the numbers back; the finished artwork is one click away in the dashboard.

• Create - short links and the QR codes that ride on them (“shorten this”, “make a QR for the spring flyer”).
• Design - save a link’s QR styling: colors, module shapes, gradients, a logo; animated designs on plans with animated export.
• Measure - totals, timeseries, countries, devices, referrers, and click quality (likely-human vs datacenter vs VPN).
• Browse - list links and campaigns, inspect one link, check the account and plan with whoami.

The server speaks JSON-RPC 2.0 over Streamable HTTP at a single endpoint on the Kilo QR origin, under /api/mcp. Most clients take a small JSON block - this is the shape (the key comes from the next section):

{
  "mcpServers": {
    "kilo-routes": {
      "type": "http",
      "url": "https://qr2r.com/api/mcp",
      "headers": { "Authorization": "Bearer $KILO_API_KEY" }
    }
  }
}

In Claude Desktop and Cursor that block goes in the MCP settings file; other clients name the fields slightly differently, but it always comes down to the URL plus the Authorization header. Claude Code wires it up in one line:

claude mcp add --transport http kilo-routes https://qr2r.com/api/mcp \
  --header "Authorization: Bearer $KILO_API_KEY"

Authentication is the same bearer API key as the REST API: mint one in the dashboard under API keys. Give the assistant a read key if it only reports numbers, or a read + write key if it should create links and designs. Key management itself stays interactive-only - no key, and no MCP tool, can mint or revoke other keys.

MCP access is included from the Pro plan up. Ask below that and the server answers with a structured error carrying the exact upgrade and pricing URLs - so an assistant relaying the failure hands you a working next step instead of a dead end:

{
  "jsonrpc": "2.0",
  "id": 1,
  "error": {
    "code": -32003,
    "message": "MCP access is included from the Pro plan up (current plan: Free). …",
    "data": {
      "feature": "mcp_access",
      "upgrade_url": "…/billing",
      "pricing_url": "…/#pricing"
    }
  }
}

The tools are deliberately phrased around the sentences people actually say. A few that work as-is:

• “Shorten https://example.com/spring-catalog and call it Spring catalog.”
• “Make a QR code for our cafe menu and style it dark green with rounded dots.”
• “How many scans did the flyer get in the last 30 days, and from which countries?”
• “Which of our active links got the most likely-human clicks this week?”

Under the hood those map to a handful of tools - create_link, save_link_qr_design, get_link_analytics, list_links - plus a whoami the assistant calls when it needs to check your plan or diagnose a permissions error. The server also publishes a getting-started resource any client can read back to you.

The tool set is read + create only, on purpose. An assistant holding your key can add things and read things; it cannot destroy or redirect anything that already exists.

• No delete, no disable, no pause - no tool removes a link.
• No destination edits - a printed QR can’t be silently re-pointed from a chat.
• Every create runs the same validation, plan caps, and destination safety scanning as the dashboard - MCP is not a side door.
• A read-only key cannot call a create tool, and nothing over MCP touches members, billing, or other keys.

Everything an assistant can do over MCP you can also do over plain HTTP - the API guides cover the same ground one endpoint at a time.