1StoreGet it on Microsoft Marketplace. Sign in with Microsoft.
2DiscoverFinds your agents from that same sign-in.
3ScanReads each agent's instructions and proposes limits, each with its quote.
4ApproveYou approve the limits. That signature is the mandate.
5GatewayOne container in your network, in front of the agent's MCP tools.
6HealthGoverned, Needs attention, Not governed: measured live.
You sign what you saw. D4A enforces what you signed.
1 of 6 · Store
Get it on Microsoft Marketplace
The question it answers: Where do I start?
Open the listing "DiaCroma for Agents (D4A)" and press Get it now. Sign in with your Microsoft work account. You land in the console with your plan active.
Open the listing on Microsoft Marketplace ↗
2 of 6 · Discover
It finds your agents
The question it answers: Which agents do I have, and what can each one call?
Nothing to paste. D4A reads your Foundry projects from your sign-in and lists every agent with its tools. It only reads: nothing in your Azure changes, nothing is governed yet.
3 of 6 · Scan
It reads what you told them
The question it answers: What limits did we already write down, without noticing?
Each scan reads one agent's own instructions and proposes limits, each with the sentence it comes from. You can leave the page; it keeps going. A proposal governs nothing until someone approves it.
4 of 6 · Approve
This is the signed mandate
The question it answers: What exactly is this agent allowed to do?
Every limit comes with the sentence it was read from. Approve as proposed, or open any agent and change it. Each approval is signed. That signed set is what D4A checks every tool call against. The screen also flags any tool whose limit the scan could not read. The first question is Enforcing or Observing: in Observing, D4A counts what it would have stopped and blocks nothing. Enforcing is the default. Connect your OPA or Cedar policies and two stricter modes appear: Strict and Explicit.†
5 of 6 · Gateway
One container, in front of the tools
The question it answers: Where is the check made, and what do I have to run?
Install once on a Linux host with Docker: one command. Point your agents at it from the console. It applies what you approved before each action, and keeps doing so if we are unreachable. In Observing, the card shows how many actions it would have stopped.
Optional · Your policies
Connect your OPA or Cedar
The question it answers: Can D4A apply the policies we already wrote?
Point D4A at your OPA or Cedar once, in your own installation, never through our panel. From then on it checks them before every action, alongside the agent’s signed limits.
6 of 6 · Health
Governed or not, measured now
The question it answers: Is it actually working, today?
Governed, Needs attention, Not governed: measured right now, not remembered. Every amber line says the cause and the button that fixes it. If your policies stop answering, Health turns red.†
Do I have to change my architecture? No. One address changes.
What stays the same
The agent: its instructions, model and runtime. The tools: the same MCP server, the same credential, which stays on your host and never reaches us.
What changes
One address: the agent's MCP tool points at D4A instead of straight at the server. The console shows what will change and writes nothing until you confirm.
What if a tool is not MCP?
It keeps working. D4A does not see it, so it says so: that route is listed as Not governed, never hidden.
No new architecture. One address changes, and it can be undone.
What our console sees, and what it never sees →
MCP: Model Context Protocol, the standard way an agent calls a tool.