The agent
engine.
Nous takes a goal, works out what to do, calls the tools that do it, and shows its work. Chat is one surface it runs on — not what it is.
Why Nous exists
An engine, not a
chatbot builder.
Most tools in this space ask you to draw a flow and hand you back a reply. Nous starts from the opposite end — a reasoning core with a tool registry, approval gates and a run log attached. The interesting question was never what it says. It is what it is allowed to do, and whether you can see afterwards what it did.
An agent, not a chatbot
Most conversational tools stop at the reply. Nous is built around the tool call — it decides what to do, does it, and the sentence it writes afterwards is a report on work already done.
Surface-agnostic
A message thread, an HTTP call, a webhook, a nightly job, a panel inside software you already run. Adding a surface is an adapter, not a second engine.
Safe by default on irreversible work
Retrieval, a tool registry, human handoff, approval gates on anything that spends money or destroys data, and a full trace of every run. The plumbing that separates an agent from a demo is in the box.
What runs on Nous
One engine,
many surfaces.
- LiveIris by CaffinixConversational commerce on Messenger — the first product on the engine
- In useOur own operations agentsFiling work, running checks and moving state across the tools we run the studio on
- AvailableYour softwareYour tools, your data, your surface — the engine embedded behind it
Under the hood
Boring stack.
Sharp choices.
We picked tools that survive — not the latest framework of the week. Each piece earns its place.
- Spring Boot
The orchestration layer and tool registry. Boring on purpose — JVM stability, clean adapters, and a runtime we are happy to leave running unattended.
- PostgreSQL + pgvector
Tenant-private retrieval and the run log. Every step, tool call and cost lands here, which is what makes a run replayable instead of merely finished.
- Gemini · Claude
Routed by step complexity. Cheap models for routine work, larger ones where judgment is genuinely needed, so cost tracks difficulty rather than volume.
Capabilities
What the engine
actually does.
Actions, not answers
The tool call is the point. The engine's job is to change something in a real system and tell you what it changed — the sentence it writes afterwards is a by-product.
Any surface
Messaging, HTTP, webhook, cron, or embedded in software you already run. A surface is an adapter; the reasoning core does not know which one it is.
Gates on irreversible work
Actions that cannot be undone stop for approval by default. An agent that can spend money or delete data without a gate is not a product, it is an incident waiting to be filed.
Every run is traceable
Each step, tool call, input, output and cost is recorded, so a run can be replayed and explained after the fact instead of argued about.
Tenant-private retrieval
pgvector memory scoped per tenant — the agent reasons over your data, and only yours.
Cost-routed reasoning
Cheap models for routine steps, larger ones where judgment is actually needed, so the economics survive volume.
Need an agent that does things,
not just says them?
Iris is the first product built on Nous, and the engine is not commerce-shaped. If you have systems an agent should be acting on — and a need to prove afterwards what it did — talk to us.