See it run
It drafted the email. It did not send it.
A support agent, asked about a refund. It searched the knowledge base, wrote a reply, and stopped — because sending mail to a customer is a medium-risk action and the policy holds those for a person. Captured from a real execution, not staged.
Experience the intelligence
Input
Retrieval
Reasoning
Governance
Action
Idle — press Run demo.
Support Agent
llama3.2:3b · local
A real run, captured from the self-hosted stack.
Captured verbatim from an execution against the self-hosted stack — the retrieval result, the drafted email and the approval summary are what the system produced. Only the timing is staged.
The retrieval is real
The passage came back from a knowledge base over pgvector, with hybrid search and a citation. Not a lookup table written for the demo.
The gate is not the model’s idea
Risk is computed from what the call would do — reversible, externally visible, what it is worth — before it runs. A model cannot argue its way past it.
It cost nothing
This ran on a local model in Docker, with no external API. The same agent runs on a frontier model by changing one setting.
Next
Find out what this is worth in your business
The demo above is one run. The scanner is about your operation — which processes are automatable, what that saves, and what it costs to get there. Four minutes, and the answer is yours either way.