Hand over the security decisions. Keep the final say.
Now in private beta. Taking early clients.
Your agents are making security decisions all day: what data to touch, what to build, what to ship. They are confident whether or not they are right, and nobody is reviewing them at the speed they run.
Omalu takes those decisions off your team. Your agents ask us what the rules are, how to meet them, and whether what they built passes. You keep the agents, the data, the execution, and the final say.
Where the line sits
Nothing moves out of your company except the questions and the answers. Your agents stay on your infrastructure, and your data stays with you.
You keep
- Your agents, running on your own infrastructure.
- Your data. We do not need a copy of it to answer.
- Execution. We decide, your systems act.
- The final say. Every verdict can be overridden by you.
- The execution log, in full.
We take on
- Which policies and regulations apply to the task.
- How the thing should be built to satisfy them.
- Whether what your agent produced actually complies.
- The judgement calls a senior consultant should be making.
One thing does not move, and we would rather say it before you ask. Accountability to your regulator stays with you. What shifts is the work of deciding, and the expertise behind each decision.
What you change: one line
You add a standing rule to your agent instructions: ask the CISO first. That is the whole integration on your side.
Everything that changes over time sits behind that first call. When a policy changes, a regulation lands, or a customer adds a contractual clause, we update it. You do not touch your system prompt again.
What is MCP?
Model Context Protocol is the standard way AI assistants and agents connect to outside tools. If your product already uses it, connecting Omalu takes an endpoint and a key. If it doesn't, we have a plain REST endpoint that does the same thing.
What your agents ask, and when
Two questions while the work is being planned, and one before anything ships.
Which policies apply?
Asked at the start of a task. Back comes what the agent must do regardless, plus the triggers that mean it has to check in again. For example, touch data of this type and you ask us before going further.
How do I build it?
Asked when a trigger fires mid-task. Back comes the approved pattern: the library to use, the key hierarchy, the vault call, a code snippet, the anti-patterns to avoid, and the policies it satisfies.
Does the result comply?
Asked before anything ships. Back comes approved, blocked, or escalated to our team, with every finding tied to a policy, the evidence for it, and a suggested fix.
What comes back
- A verdict: approved, blocked, or escalated to a person.
- The reason, in language a non-engineer can read.
- The specific control, regulation or standard it maps to.
- A timestamped record of the question and the answer, exportable for an auditor.
Why not just ask your own model
Your own model will answer confidently either way, and it is grading its own homework. We check the artefact itself against the policy set, in a clean context. We do not read the agent's account of what it built, because that is the thing under review.
What you are buying is an independent second opinion, backed by a named, CISSP-certified security professional who stands behind the methodology, and an evidence trail that does not come from the system being checked. When a customer's security team asks how you control what your agents do, "we built our own prompt for it" is a harder answer than this one.
How it runs
- Human escalation. Reviews that come back uncertain, or that cross a risk threshold you set, are routed to a senior consultant and answered within one business day. You see which answers were automated and which were reviewed.
- Integration. MCP server endpoint plus an API key. Works with Claude, Claude Code, and any MCP-compatible agent framework. REST available if you're not on MCP.
- The log is yours. Every question, every verdict, and the reason behind it, kept on your side and exportable whenever you want it.
Questions
What is MCP and do I need it?
MCP is the standard connection between AI assistants and outside tools. You only need it if you're building an AI product of your own. If you're not, this page isn't for you.
We're building AI agents. Isn't this what guardrails are for?
Guardrails stop the obvious. This answers the judgement calls: the architecture, the compliance question, the action that's technically allowed but a bad idea. And it leaves a record that you asked.
Request access
Tell us what you're building and how your agents are making decisions today. Thirty minutes, no pitch. If this isn't the right fit, we'll tell you that too.
We reply to every enquiry within one business day.
Not building AI agents? See the other two ways to work with us →