Vivari Guard · Updated July 2026
The senior engineer that
never sleeps, never guesses.
AI agents ship changes faster than humans can review them — and reviewing an LLM with another LLM inherits the same blind spots. Vivari Guard is the deterministic safety layer: evidence from your own git history, a reproducible risk score, and a gate that holds what deserves holding. No LLM in the loop.
Evidence the call graph cannot see
Every codebase has invisible couplings: the config that must track the schema, the fixture that always moves with the parser, the override in a file no import connects. Static analysis can't see them — your git history can. Guard mines it into persisted change-coupling facts: "these two files changed together in 14 of 15 commits — and this diff touches only one of them." Every finding is a fraction, a date range, and named commits. Not a model's opinion — a fact from your own history, checkable by anyone.
A risk score you can put in front of an auditor
Each change is scored 0–10 across 17 weighted signals — contract propagation to un-updated callers, blast radius, hub criticality, historical churn, test coverage, author familiarity, and more. Signals with no real data are excluded and their weight redistributed, never silently zeroed. The same diff always produces the same score. Deterministic, traceable, reproducible — the properties compliance teams actually require from AI agent governance, and the properties "an LLM said it looks risky" can never have.
PASS / WARN / BLOCK — where the work happens
Guard sits in the working loop, not in a dashboard nobody opens: a pre-commit hook and a CI gate that returns PASS, WARN, or BLOCK with the evidence attached — as terminal output, PR comments, or SARIF for your existing code-scanning pipeline. Analysis completes in under a second with no network call. Inside the Vivari workspace, every hold an agent triggers becomes a one-tap human decision with the evidence in view — and every decision lands on the audit trail, replayable and exportable.
Local-first, by architecture
Your code never leaves your machines. Guard indexes and analyzes on your own infrastructure — seven languages, hundreds of thousands of symbols, cold-rebuild in seconds — which is why it can gate every commit without a privacy conversation.
Guard ships as a layer of Vivari — and a standalone offering for teams that want the safety layer first is planned. Request early access and mention Guard; standalone interest shapes the rollout.
Request early accessWhere Guard fits in agent management →