vantio-open-core · Vantio Optics

Vantio Optics — See LLM egress. Blind by design, not a proxy.
The Free · Open Core observe plane: intercept outbound LLM calls without code changes, capture metadata (never prompts), and export proof — observe only, no enforce.

This repo ships Vantio Optics. The same client runs in every tier; higher tiers unlock Vantio Gate (Pro · enforce) and Vantio Phantom Engine (Enterprise · absolute control) on top of Optics.

Vantio tiers unlock governance:
Vantio Optics (Free · Open Core) → Vantio Gate (Pro) → Vantio Phantom Engine (Enterprise)

Get started in 60 seconds

npm install -g @vantio/cli
vantio run node agent.js          # no key needed — Free Observe works immediately

No-install path:

npx @vantio/cli run node agent.js

Optionally connect to your dashboard (Pro/Enterprise features):

vantio login <your-api-key>       # validates + saves the key; no env vars needed

Grab your API key from vantio.ai/dashboard.

After a run, generate an auditor-ready proof artifact or explore local history:

vantio prove                      # HTML proof artifact (Free, no key needed)
vantio discover --local           # local run history (Free, no key needed)
vantio discover                   # full workspace history (Pro/Enterprise)

Sight Loop (Optics workflow)

Optics ships one named workflow — Sight Loop:

1. Wrap — vantio run / SDK
2. Capture — host · process · bytes · time · trace (no prompts/completions)
3. Inspect / export — vantio prove, vantio discover --local, or Optics MCP
4. Honest residual — ungoverned paths stay silent → upgrade cue

Full walkthrough: docs/sight-loop.md · MCP: docs/optics-mcp.md · Offline prove: ./scripts/sight-loop-prove.sh

Docs

Packages

Tier model — one platform, unlocked layers

Optics is the client at every tier. On Free (no Pro key), it observes only — events are labelled OBSERVED. With a Pro key, the same client fetches policy from Vantio Gate and can block, redact, or cap locally. Enterprise adds Vantio Phantom Engine beneath the app layer. See observe-only.md for the Free-tier fence.

Full breakdown: PRODUCT_LINEUP.md

SDK — explicit trace correlation

Node.js

npm install @vantio/agent-sdk

import { shield } from "@vantio/agent-sdk";

await shield(async () => {
  await runMyLLMAgent();
});

Python

pip install vantio-agent-sdk

On Ubuntu/Debian (23.04+), global pip install is blocked by default (PEP 668). Use a virtualenv or pipx install vantio-agent-sdk instead.

from vantio import shield

@shield
async def run_agent():
    await call_openai(prompt)

Supported LLM providers

api.openai.com · api.anthropic.com · generativelanguage.googleapis.com · api.cohere.ai · api.mistral.ai · api.groq.com · api.together.xyz · api.perplexity.ai · inference.ai.azure.com

What gets captured

• Which LLM endpoint was called
• Response size in bytes
• The process ID
• A trace ID that links calls across your agent's full execution

What never gets captured: prompts, completions, or any content from your requests.

Optics: honest about bypassability

vantio run intercepts LLM calls by patching globalThis.fetch in the Node process. This covers the vast majority of agents without code changes.

It can be bypassed — by native socket calls, un-instrumented subprocesses, or processes not started with vantio run. This is intentional. Optics surfaces your governance gap; it does not paper over it.

• Vantio Gate (Pro) — Policy Latch: block, redact, caps, dashboard sync, fleet discovery. App layer.
• Vantio Phantom Engine (Enterprise) — Ring-0 eBPF TLS observe, Bypass Reconciliation when app and kernel diverge.

Use vantio discover --local to see what Free observes on your machine. Use vantio prove to generate an auditor-ready proof artifact from any run.

Telemetry

Anonymous, opt-out usage analytics. No prompts, completions, API keys, or emails. Opt out at any time:

export VANTIO_TELEMETRY_DISABLED=1   # or
export DO_NOT_TRACK=1

vantio.ai · Dashboard · Pricing · MIT License