Vantio Optics MCP server

Vantio Optics (Open Core)

Free, local-first observability for AI agent egress. See what your agents call out to — without ever reading the conversation.

npm install -g @vantio/cli
vantio run node agent.js

That's the integration. Optics intercepts outbound calls to known LLM providers and reports destination, process, size, and timing — never prompts or completions.

Python: pip install vantio-agent-sdk, then wrap your agent with @shield.

Optics is the free Observe tier. Phantom Engine (Enforce + Control) extends this to policy enforcement and host-level runtime protection on enrolled Linux systems. Full docs: vantio.ai/docs.

Supported runtimes

That's the only install step. Works on macOS, Linux, and Windows (WSL).

On Node: fetch, undici, http/https, http2, net/tls, WebSocket, and spawned curl or wget.

vantio run only injects the Node interceptor (NODE_OPTIONS --require) for node / npx / tsx / ts-node.

Python support requires vantio-agent-sdk. Follow the current Python SDK example and verify that a supported outbound event appears before relying on the coverage state. Prefixing vantio run python does not intercept by itself.

Browser paths stay outside this wrap.

What is recorded

Optics records connection and process metadata from supported wrapped agent paths. Does not block actions or retain prompts or completions.

Vantio records that a call was made, when, to which provider, and how many bytes — nothing more.

What is never recorded

• The content of your prompts
• Model completions or responses
• Any personally identifiable information

Present coverage boundaries

vantio run intercepts LLM calls by patching globalThis.fetch in the Node.js runtime (via NODE_OPTIONS --require). This covers the vast majority of real agents — every OpenAI SDK call, every LangChain.js call, every Vercel AI SDK call — without any code changes.

It can be bypassed. A process can:

• Call an LLM endpoint directly through a native socket without using fetch
• Spawn a subprocess that isn't started with vantio run
• Use a language runtime other than Node without installing the Python SDK

This is intentional, not a bug. Vantio Optics surfaces your governance gap honestly.

Without the SDK, vantio run python does not intercept. Browser paths, or processes not started with vantio run / shield(), never hit Optics — and Optics does not invent a record for them.

License and telemetry

MIT License.

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

Upgrade path to Phantom Engine and Enterprise

Free Optics needs no account and no API key. vantio login is optional and later, for dashboard sync only, never a required step before running.

Optics is the free Observe tier. Phantom Engine (Enforce + Control) extends this to policy enforcement and host-level runtime protection on enrolled Linux systems.

Residual risk closes with Vantio Phantom Engine ($799/node/mo) — see vantio.ai/pricing. Enterprise governance is available on top — talk to sales.