Ontology / policy language
Typed meaning and allowed actions — what agents may mean and do.
Agent verification infrastructure
Define what your agents may mean and do as an ontology. We compile it into a gate on every LLM tool call — with a readable proof of why it passed, failed, or needs a human.
The problem
Production agents call tools that charge cards, mutate systems of record, and message customers. Teams either keep agents in a sandbox forever or ship and hope.
There is no compile-to-gate path — until now.
How it works
Velstack is agent automation and verification infrastructure. You write policy once. We sit in front of any tool call and return a decision you can audit.
Policy as ontology: entities, allowed actions, constraints, evidence requirements.
Verification compiler turns policy into runtime gate artifacts.
SDK wraps each tool call → pass / fail / needs-human + proof trace.
Product surfaces
Typed meaning and allowed actions — what agents may mean and do.
Policy → gate artifacts. Versioned, testable, reproducible.
Wrap tool calls; get pass / fail / needs-human + proof.
Versions, traces, human-approval queues, audit export.
Who it's for
Buying trigger: the first tool call that can charge, refund, book, delete, escalate, or send — and someone asks how do we prove that was allowed?
Mechanized trust for agents other people already run — not another agent. Not prompt-only governance. A gate on the call path with evidence.
Proof trace
Mini demo: a refund tool call fails without manager approval, then passes when approval evidence is present — with a readable proof of which rule fired.
Get started
Start with one gated tool call. Design partners welcome — early stage, building with teams who feel the blast radius.