Every AI change, proven.
Use AI on your regulated code without losing the proof. Anchored traces every change your AI coding agent makes to the requirement it implements, the test that proves it and the person who signed it off, and keeps the record your auditor asks for current.
Built for medical-device software under IEC 62304 and EU MDR. Runs inside your repository and your build, with Claude Code.
They give the AI the rules. We make it follow them.
Requirements tools now hand your specification to the AI as context. Context can be skipped. Anchored sits in the agent's working loop and in your build, so a broken rule is stopped before it goes any further.
- The agent types a number from the specification into the code Stopped on that file. The agent is told which requirement the number belongs to and where to read it from. Stopped
- The code cites a requirement that does not exist Stopped, with the unknown ID named. No code can claim to follow a rule nobody wrote. Stopped
- A requirement nobody has approved yet The agent will not start building on it, and names the person who has to approve it. Waits for a person
- A requirement without a passing test named after it The build fails. A comment that claims coverage is not evidence; a test is. Build fails
The record your auditor asks for, always current.
Every requirement, where it comes from, who approved it, the code and the test that proves it. Generated from the same data the checks use, stamped with the code version, exported as CSV or a printable page.
| Requirement | What it says | Source | Approved by | Code | Test | Status |
|---|---|---|---|---|---|---|
| GA-LOW-001 | Alert when a reading is below 3.9 mmol/L | §4.2.1 | A. IvanovaClinical lead, 6 Oct 2026 | evaluate.ts:21 | evaluate.test.ts:7 | Proved |
| GA-LOW-002 | Urgent alert below 3.0 mmol/L, cannot be snoozed | §4.2.2 | A. IvanovaClinical lead, 6 Oct 2026 | evaluate.ts:22 | evaluate.test.ts:12 | Proved |
| GA-DATA-001 | Never raise or clear an alert from a reading older than 15 minutes | §4.4.1 | A. IvanovaClinical lead, 6 Oct 2026 | evaluate.ts:16 | evaluate.test.ts:16 | Proved |
| GA-UNIT-001 | Convert mg/dL to mmol/L before any comparison | §3.1 | A. IvanovaClinical lead, 6 Oct 2026 | evaluate.ts:11 | evaluate.test.ts:22 | Proved |
From the glucose-alert example that ships with Anchored; the specification and the approver are illustrative. The IEC 62304 export adds software safety class and risk controls per requirement.
Four steps, one of them human.
We call the method Anchor-based development. The part of your code a specification governs, usually 15 to 25 percent, is anchored to it; everything else is built the way you build it today.
Bring in your specification
The agent drafts one numbered requirement per rule, each pointing to the paragraph it came from, all marked unapproved, with every unclear point listed.
A person approves each one
Your clinical or regulatory lead decides what each requirement means. Only the people you name can approve.
Every change is checked
As the agent writes, and again in your build. Each requirement needs its own passing test.
Export the proof
One command produces the traceability record for the current code version, ready for the audit.
What it does not do.
Your quality team and your notified body decide compliance. Anchored gives them evidence they can check.
It does not interpret your rules
If your expert approves a reading, the code follows that reading, and the record shows who approved it.
It does not make you compliant
It proves that the code does what the approved requirements say. It is not legal or regulatory advice.
It does not take your code anywhere
The checks are plain Python scripts that run in your repository and your build. Nothing is sent to us.
Anchored for IEC 62304.
Start with a 20-minute call: you tell us how you keep traceability today, we show Anchored on the glucose example. If it doesn't fit, we say so.
For the first five teams
Design Partner
€2,500 one time
- We set it up with you, in your repository
- Your requirements, checks and first audit export, done
- Three months included
- In return: your feedback and a short case study
Set it up yourself
Traceability
€490 per repository, per month
- Installs in two commands
- IEC 62304 export included
- Updates and email support
- Cancel any time
You pay only once we agree it fits. Prices exclude VAT.
You talk to the person who built it.
I set up every design partner myself, in your repository and with your team, and stay with you for the three months. The code is plain Python you can read and keep.
Stoyan Dimitrov, newcleardigital
Questions.
We already use Jama, Polarion or Ketryx.
Keep them. They hold the requirements; Anchored works where the code is written and stops untraced changes before they exist. Its evidence exports as a CSV you can bring into those tools.
We don't use AI on certified code yet.
Then this is the easiest time to start: setup is simplest before the first AI-written line, and the record is complete from day one.
Is it only for medical devices?
No. It fits any software that must follow a written document: payment and banking rules, insurance tariffs, safety standards for cars, trains and aircraft, or a contract. Medical-device software is where we start.
What do we need?
Claude Code and a GitHub account. The checks work with TypeScript, Python, Java, Kotlin, C#, Go, Rust and more.
Does our code leave our servers?
No. The checks run in your repository and your build. Nothing is sent to us.