We let Claude make our QR codes for a week. Here's what happened.

Quick context: qrlia is the QR-code platform Castle Group has been operating internally for ~18 months. We launched the public version this month. One of the bets we made early was that the Model Context Protocol (MCP) — Anthropic's spec for letting LLMs talk to your tools natively — was going to become table stakes for B2B SaaS by mid-2026. So we shipped MCP support on day one.

I wanted to know if the bet was real. So I did the most credible test I could think of: I gave Claude Desktop a tenant API key and asked it to run my QR fleet for a week.

The setup

Six lines of config in ~/.claude/mcp.json:

{
  "mcpServers": {
    "qrlia": {
      "type": "http",
      "url": "https://mcp.qrlia.com/mcp",
      "headers": { "Authorization": "Bearer qrc_<your-tenant-key>" }
    }
  }
}

Restart Claude Desktop. /mcp shows qrlia connected, ~36 tools available — qrs_create, designs_create, scans_aggregate, the lot.

Then I asked Claude things in plain English. No fine-tuning. No prompts beyond what you'd say to a colleague.

What it actually built

> "Mint a vCard QR for our new sales rep, Sara. Her email's sara@castlesoft.eu, phone +971 50 123 4567, title 'Account Executive'. Brand it with the Castle Group blue."

Claude pulled the tenant theme, populated the vCard, picked the right module shape to match the existing brand kit, ran the design through qrs_create, returned a qrlia.com/r/sara-castle short link. 18 seconds.

> "We're running an event in Dubai next Wednesday. Spin up 12 booth QRs labelled Booth-1 through Booth-12, all pointing at the event's multilink page. Use the silhouette of our logo for the QR shape."

This was the test. The silhouette mask is a Pro feature; Claude had to (a) call qrs_create 12 times, (b) attach the silhouette via a separate designs_create, (c) handle the rate limiter when our tier limits hit. It did all three. 4 minutes. The output was 12 logo-shaped QRs that scan, named consistently, all in the right folder.

> "Show me which of yesterday's scans came from outside the UAE."

scans_list with a date filter, then a JOIN against the geo-rollup. Claude wrote the query as 3 successive tool calls, returned a markdown table with country breakdowns and a sparkline. Under a minute.

Where it broke

Twice. Both my fault.

1. I asked it to "delete the QRs from last month's campaign" with no further specification. It refused — wanted me to either name the QRs or confirm a folder. Right call. Destructive operations on a populated tenant should never happen on a vibe.

2. I asked for "a QR that loops between three URLs randomly." This isn't a feature qrlia ships — we have geo + device + schedule rules, but no random splitter. Claude tried to fake it by making three separate QRs with a random redirect, then realised there was no random redirect. It surfaced "this isn't a thing your platform does — want me to set up a 50/50 A/B between two of them instead?" That's the right shape for a tool that hasn't been over-promised.

The week's stats

• 47 QRs created by Claude end-to-end (vs my baseline of about 15/week clicking through the dashboard).
• 3 scan-analytics reports turned around in under 2 minutes each — work that previously meant me opening a spreadsheet.
• Zero misfires that hit production. The two failures above were caught before any QR went out the door.
• Zero finetuning — Claude is reading the tool schemas qrlia ships and figuring it out from natural language.

Why this matters

I think the model of "your AI assistant has a bearer key for your B2B tools" is going to be normalised by Q4 2026. Consumers are already used to Siri / Google Assistant doing simple delegations on their behalf. The B2B equivalent — "Claude, mint the vCard for my new hire" — is just the same idea wearing a tie.

If you're running QRs at any volume — events, real estate signage, restaurant menus, sales-ops vCards — and you're already paying for Claude or Cursor, the marginal cost of adding qrlia to the conversation is six lines of config.

And if you're building a B2B SaaS yourself: ship MCP. Don't wait for the consultant.

Try the qrlia MCP gateway →