How Do You Check an AI Agent's Medication Claims Against the FDA Label?

Originally published at orinyx.io.

Not for clinical use. Orinyx checks medication claims against cited FDA label text for testing and evaluation only. Do not use its results to diagnose, treat, prescribe or make patient care decisions. Use synthetic test data only.

**The problem
**An agent that writes medication orders can state a dose, route or frequency the FDA label doesn't support, and sound just as sure as when it's right. The usual fix is to ask a second model to grade the first. But research finds that language-model evaluators favor their own outputs, so the grader can approve a convincing wrong answer. Your test suite then passes for the wrong reason.

**What Orinyx does about it
**The Orinyx MCP is a free, open-source server your agent calls like any other tool. It checks each dose, route and frequency against FDA label text (read through openFDA) using fixed rules, not a model. Every result cites the label section it used and links to the DailyMed page, so a person can open it and read it.

Each claim comes back as one of four answers:

• Supported: consistent with the cited label text.
• Flagged: the label says otherwise, with the section cited.
• Unreadable: the order can't be read as written.
• Not verified: Orinyx couldn't check it. This never means the claim is correct.

There is no "safe" answer, and a run with zero flags is not a clearance.

**Why 0.1.4, and what it changes for you
**Version 0.1.4 is a documentation release. The hosted service already let developer accounts use the five FDA reference lookups, but the open-source README, changelog and tests said otherwise. I fixed the docs to match. No tool name, input schema or verdict changed, so nothing you've built breaks.

What this means in practice:

• Repeatable tests. The same input gives the same answer, so a failing check is a real finding you can reproduce, not a model's mood.
• Honest coverage. "Not verified" is its own result, so a gap in your agent's checking shows up instead of passing silently.
• Docs you can trust. What the README says is what the service does.
• Runs in CI. The package includes a REST client and a CI recipe, so you can check medication claims on every change.

**Try it
**Sign up free (200 checks a month), then connect the hosted connector at https://orinyx-mcp.fly.dev/mcp or run the server locally. Claude Code is the client I've tested; other routes are documented in the README. Use synthetic data only.

Known gaps are listed in the known limits, and the full explainer is on orinyx.io.