Two Things Claude Already Catches That We Never Told You About

15 Jul 2026

15 Jul 2026 by Luke Puplett - Founder

We're bad at telling people what we've shipped. Two examples turned up this week that are worth an exception — both already live, both just sitting in the docs where nobody but a developer would stumble on them.

Both come from the same place: connect Claude Desktop to Zipwire over MCP, and it doesn't just fetch your data. It reads it.

It notices when a timesheet doesn't add up

Every submitted timesheet gets scanned for a set of conditions worth an approver's attention — worked a public holiday, worked a weekend, billed more or fewer hours than a standard day, no rate plan in effect. These show up as flags: short banners on the timesheet, information rather than a verdict.

Ask Claude to help process a timesheet, and it doesn't just relay the numbers — it reads the flags and explains what they mean.

In this example, nobody asked Claude to check anything. It fetched a timesheet for processing, noticed the period predates the sender's rate plan — which is why the charge showed £0 — flagged Tuesday's 13 hours against an 8-hour standard day, spotted the underworked days either side, and then offered to fix the rate plan backdating before anyone actioned it.

That's the gap between a tool that retrieves data and one that understands it.

It catches what the green checkmarks don't

The second one is sharper. Right to Work checks in Zipwire run through Yoti, a certified identity service provider — document authenticity, liveness, face match. When all of those pass, the card shows green across the board.

Green checkmarks confirm the passport is genuine and the selfie matches the passport. They don't confirm the passport belongs to the person who submitted it — because that's not what they're checking.

Here, every individual Yoti check passed. But Claude compared the self-declared identity on the submission against the actual passport data Yoti's report had read — and they didn't match. Different name, different date of birth. It recommended rejecting or querying the submission rather than accepting it on the system's default, before a human had the chance to skip past a card that looked entirely green.

That's the Collect → Yoti ID check → Claude pipeline in practice: Zipwire gathers the evidence and runs the automated checks, Claude adds a layer of scrutiny on top by reading the underlying data rather than trusting the pass/fail badges — the kind of cross-check that's easy for a person to miss when every box looks fine.

Why this is worth saying out loud

Neither of these is a new feature we built for this post. They're just what happens when you give an AI assistant read access to real records instead of a dashboard summary. A dashboard shows you a green tick. Claude can look at what's behind the tick.

Worth being precise about what this isn't: Claude isn't a replacement for the ID check itself, and it isn't infallible either — treat its read as a second pair of eyes, not a final verdict. The value here is in surfacing a discrepancy a person might not think to go looking for, not in removing the person from the loop.

Try it yourself

If you're already on Claude Desktop, add Zipwire as an MCP connector at https://zipwire.io/api/v1/mcp and ask it to show you timesheets with flags, or pull up a Right to Work collection mid-review. Full setup steps are in the Using Zipwire with Claude guide, which is also where these two screenshots live if you want the fuller writeup.

For the broader idea of proof you can check without trusting a dashboard, see Verifiable Timesheets and AI Gates: Privacy-Preserving Compliance.

---

Learn more