Security posture
Security
Vivari runs AI agents against real codebases, so its security model is the product. The design is local-first: your repositories are never uploaded to us, each agent is scoped to its role, and everything it does lands on a record you can export. Here is where it stands today.
Local-first architecture
The workspace runs on your infrastructure, and Vivari Guard, the deterministic review engine, indexes and analyzes your repositories in place. Guard's change-safety analysis (change-coupling from your git history, the risk score, the PASS/WARN/BLOCK gate) completes with no network call. Model inference is separate: cloud models see what your harness sends them under that provider's terms, and fully local models close even that loop on your hardware. Your source is not uploaded to us to be reviewed.
Scoped, per-agent permissions
Every agent runs under a role scoped to its niche: least privilege by default. A reviewer that cannot write; a writer that cannot deploy. Risky tool calls hit an explicit gate you answer in one tap: run once, add to an allowlist, or reject. Nothing holds the production keys because it happens to be running.
Deterministic review
Guard is deterministic: the same diff always produces the same score, and every finding is traceable to named commits and measurable signals. No LLM makes the safety call, which keeps a model's blind spots out of the loop and means a review decision can be replayed when someone asks how it was made.
Full audit trail, exportable
Every tool call, message, decision, and permission grant is captured on a timestamped, replayable audit trail, exportable to your SIEM. When an agent does something you need to account for, the trail turns "we think it did X" into exactly what it did, and when.
Private repositories
Vivari points your existing agents at your existing repositories. It does not require you to push code to a third-party host, mirror it to us, or make anything public to be managed.
No training on your code
We do not train models on customer code, prompts, transcripts, or repository contents. Vivari runs the agent harnesses you already trust (Claude Code, Codex, Cursor); their own providers' data terms apply to those sessions, and we add nothing that repurposes your work.
What we do not claim
Vivari is pre-launch and holds no formal security certifications today: no SOC 2, no ISO 27001. Formal audits are on the roadmap as the product and team grow. If certification is a blocker for your organization, we would love to hear from you.
Frequently asked questions
- Does my code leave my machines?
- Vivari never uploads your repositories to us, and Guard's analysis makes no network call. Cloud model inference sees what your harness sends it, under that provider's terms; fully local models keep everything on your infrastructure.
- Do you train on my code or data?
- No. We do not train models on customer code, prompts, transcripts, or repository contents.
- Are you SOC 2 certified?
- Not yet. We hold no formal certifications today. Email hello@vivari.ai to talk through your requirements.
- How do you limit what an agent can do?
- Per-agent roles scoped to the niche, plus an explicit gate on every risky tool call.
- Can I get an audit trail for compliance?
- Yes. A complete, timestamped, replayable trail with SIEM export, plus deterministic Guard verdicts.
Security questions: hello@vivari.ai. That address reaches the people building the product.
Request early accessHow Guard reviews changes →