tria-ai

Python SDK for Tria intent-aware runtime governance.

0.5.0a1 is an alpha release. Start in audit mode, validate decisions, then enable enforcement for selected agents in the Tria dashboard.

Install

python -m pip install --pre tria-ai

Connected quick start

from openai import OpenAI
from tria import TriaClient
from tria.integrations.openai import wrap_openai

tria = TriaClient(
    api_key="YOUR_TRIA_API_KEY",
    backend_url="https://tria.intellectmachines.com",
    mode="audit",
)

intent = tria.register_intent(
    agent_id="support-bot",
    natural_language="Answer customer questions using support documents",
    permitted_actions=["lookup_order", "search_knowledge_base"],
    excluded_actions=["delete_order", "modify_payment"],
)

client = wrap_openai(
    OpenAI(api_key="YOUR_OPENAI_API_KEY"),
    tria=tria,
    intent_id=intent.intent_id,
)

register_intent() writes the intent to the backend and stores the same backend-issued ID locally. It raises if connected registration fails; it never silently creates a local-only policy.

Enforcement

Set the organization or agent to enforce in Tria before using mode="enforce". The SDK fails closed when an enforce-mode client reaches an audit-only backend. Unknown tools are sent for policy evaluation instead of being silently allowlisted.

Local-only mode

Omit backend_url to use local intent evaluation and deterministic guards. Local decisions are not persisted or cryptographically signed by the Tria control plane.

Links

• Website: https://intellectmachines.com/
• Dashboard: https://tria.intellectmachines.com/ui/
• Documentation: https://intellectmachines.com/docs.html
• Source: https://github.com/IntellectMachines/tria-sdk

Apache-2.0