vantio-agent-sdk · Vantio Optics (Python)

Vantio is the infrastructure control layer for autonomous AI. This package is Vantio Optics Observe for Python: record where the agent calls LLM APIs (host, size, process, trace) without storing prompts or completions. Optics is free. Optics does not block, redact, or cap spend on its own.

Optics helps you see. Phantom Engine is runtime protection on enrolled Linux — Observe, Enforce, and Control in one purchase. Contact Vantio. Enterprise adds governance when you need proof and process on top. Contact Vantio. Continuous Assurance is how the suite stays true after install; it is not a separate product.

Optics (Free) · Phantom Engine (Contact Vantio) · Enterprise (Contact Vantio)

pip install vantio-agent-sdk

Prefixing vantio run python without this SDK does not intercept. 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.

pip install vantio-agent-sdk
vantio run python agent.py

Optics: vantio.ai/optics · Pricing: vantio.ai/pricing · Docs: vantio.ai/docs

3.1.0 — HTTP outcome and customer lines

A stored HTTP call sets ok from the status code. ok is true for 200–399 and false for 400–599. A 4xx or 5xx response is the provider's outcome. It is not stored as network_error. DNS, a refused or unreachable connection, TLS, and a timeout still use network_error, and they have no HTTP status. A wrapped application exception is stored as the exception class name only, with no HTTP status and without network_error.

Those transport cases are recognized on urllib, requests, httpx (sync and async), aiohttp, and urllib3. Classification walks a bounded exception chain: __cause__, __context__, a .reason that is itself an exception, and exception args. It does not walk arbitrary object attributes, and it does not reclassify an exception from words in its message. The depth and visit caps are an internal safety limit, not a compatibility promise. A final HTTP status on an attempt wins over a transport error nested on that same attempt. A retry that the client handles inside one call is recorded as the final response. A call Optics observes on its own stays its own record.

Each recorded call keeps two machine fields and three customer lines:

APPLICATION_ERROR stays the machine category for an unsuccessful provider response. The customer line is the observed outcome for that status, not a generic error label. The run log names the headings in status_labels and sets schema_status to unstable-pre-1.0. Mixed machine outcomes in one run are PARTIAL. The summary line for that run is Partial.

When the hostname is in the supported provider catalog, the record names that provider and the destination, and transport lines say Provider. Any other host, including an extra host, is an upstream service: the response heading is Upstream response, and transport lines say Upstream service (for example Upstream service could not be resolved, Connection to upstream service failed, Secure connection to upstream service failed, Upstream request timed out). HTTP status lines still name the observed result, such as Provider authentication failed for HTTP 401. Optics does not block, reject, or enforce the provider result.

http.client and pycurl still do not store an HTTP status on the success path. That outcome stays UNAVAILABLE, with the line Provider outcome unavailable and the response No HTTP response.

A trace with no supported call does not write a run log. The empty view is Optics status Not observed and observed outcome No supported AI call observed.

Telemetry is unchanged: it stays off unless VANTIO_TELEMETRY=1. VANTIO_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 still override that opt-in. Phantom Engine and Enterprise APIs stay separately provisioned. Free Optics still needs no account and no API key.

3.0.15 — telemetry wording

Anonymous usage telemetry stays off unless you set VANTIO_TELEMETRY=1. The wire payload is unchanged from 3.0.14. Free Optics still needs no account and no API key.

3.0.14 — packaging metadata only

This patch corrected PyPI project URLs and the long description to the public ladder. It does not market another standalone public SKU. SDK behavior is unchanged from 3.0.13.

v3.0.x — Breaking change from v2.x

v3.0.0+ is a complete rewrite. The old VantioSession / VANTIO_PROXY_ENDPOINT API is removed. The new API uses @shield (a decorator or async context manager) with zero dependencies and no proxy — observe runs inside your SDK, not through an external endpoint.

Migrate from v2.x:

# Old (v2.x) — remove this
from vantio.session import VantioSession
with VantioSession(agent_name="my-agent") as session: ...

# New (v3.x) — use this instead
from vantio import shield
@shield
async def run_agent(): ...
# or: async with shield() as ctx: ...

shield() — trace context

from vantio import shield, report_anomaly

# Decorator form
@shield
async def run_agent():
    result = await call_openai(prompt)
    return result

# Context manager form
async with shield() as ctx:
    print(f"Trace ID: {ctx.trace_id}")
    result = await run_agent()

get_current_trace_id()

Returns the active trace ID for the current async context, or None outside a shield() frame.

from vantio import shield, get_current_trace_id

async with shield() as ctx:
    trace_id = get_current_trace_id()  # same as ctx.trace_id

get_current_trace_id()  # None — outside shield() frame

fetch_policy() — policy retrieval

Scope: this API is part of Vantio Phantom Engine / Enterprise and requires a separately provisioned control-plane key. It is not part of free Vantio Optics, which runs local-first with no account and no API key.

Fetches your tenant's governance policy from GET /api/v1/config. Fails open — if the control plane is unreachable, a permissive default policy is returned so your agent never stalls waiting for governance.

from vantio import fetch_policy, redact_pii
import os

policy = fetch_policy(os.environ["VANTIO_API_KEY"])

# Check enforcement flags
if policy.enforce:
    if "api.openai.com" in policy.blocked_hosts:
        raise RuntimeError("OpenAI is blocked by policy")

# Use with redact_pii for SDK-side PII scrubbing
if policy.pii_redact:
    result = redact_pii(user_input, policy.pii_types)
    prompt = result.text   # PII scrubbed before the LLM call

VantioPolicy fields:

redact_pii() — local PII scrubbing

Pure, side-effect-free local PII redaction using the same patterns as the CLI interceptor and @vantio/agent-sdk. No content leaves the process.

from vantio import redact_pii

result = redact_pii("Contact bob@example.com or call 555-123-4567")
# result.text       → "Contact [VANTIO_REDACTED:EMAIL] or call [VANTIO_REDACTED:PHONE]"
# result.redactions → ["email", "phone"]

Pass a custom pii_types list to control which categories are scanned (values are case-insensitive):

result = redact_pii(text, pii_types=["email", "ssn"])

Built-in categories: ssn, email, credit_card, phone.

RedactionResult fields:

report_anomaly() — cloud ingest

Scope: this API is part of Vantio Phantom Engine / Enterprise and requires a separately provisioned control-plane key. It is not part of free Vantio Optics, which runs local-first with no account and no API key.

async with shield():
    await run_agent()
    await report_anomaly(
        target_host="api.openai.com",
        bytes_severed=14382,
        # Valid values: "OBSERVED" | "ALLOWED" | "REDACTED" | "BLOCKED_HOST" | "BLOCKED_SIZE" | "BLOCKED_SPEND"
        action_taken="BLOCKED_HOST",
    )

report_anomaly() must be called within a shield() context. It is a no-op unless VANTIO_CLOUD_INGEST=true. Non-fatal — never crashes the agent.

Environment variables

Free Optics runs on this machine with no account and no API key. These three variables only control the optional anonymous ping:

The variables below are for a separately provisioned Phantom Engine or Enterprise control plane. They are not required for free Optics.

Telemetry

Telemetry is disabled by default. Set VANTIO_TELEMETRY=1 to opt in. VANTIO_TELEMETRY_DISABLED=1 or DO_NOT_TRACK=1 override that opt-in.

The automatic ping runs once per process, on the first shield() entry, before HTTP observations are recorded. On that ping, callCount is 0 and hosts is empty. Calling shield() again in the same process does not send a second ping.

When a ping is sent, the body contains anonymousId, runtime, runtimeVersion, os, event, hosts, and callCount. hosts is a list of LLM hostnames, at most 50. Optional fields, only when a caller sets them, are sdkVersion, redactedCount, blockedCount, and framework. The ping does not include prompts, completions, API keys, or email.

What gets captured

The local Optics record stores:

• Which LLM endpoint was called
• Response size in bytes
• Process ID and timestamp
• A trace ID linking calls in the same shield() / vantio run python wrap

That local record is not the telemetry ping. The ping's fields are listed in the telemetry section above.

Prompts, completions, and request content are never stored.

Zero dependencies

Core tracing requires only the Python standard library (contextvars, asyncio, hashlib, hmac, re). Cloud ingest and anonymous telemetry use urllib.request and threading. No aiohttp, no httpx, no requests.

MIT License · vantio.ai/optics · vantio.ai/phantom · vantio.ai/enterprise · vantio.ai/pricing · vantio.ai/docs