Integrator
“I want to call agents from my application.”
Python SDK or HTTP API. You submit an execution, track it, and check whether the action actually happened.
Three separate paths. Each answers a different question.
Integrator
“I want to call agents from my application.”
Python SDK or HTTP API. You submit an execution, track it, and check whether the action actually happened.
Connector developer
“I want to plug my system into the AI.”
You write a declarative manifest, verify it locally and get certified. No platform code is written.
Pack author
“I want to package an entire vertical.”
A Domain Pack holds business operations, verification rules and connector requirements in one file.
Get your key — put it in DAVIRIX_KEY.
Install the SDK:
pip install davirixSend your first execution:
from davirix import Davirix
dx = Davirix(api_key=os.environ["DAVIRIX_KEY"])n = dx.run(agent_id="card-support", tenant_id="bank-uz", input={"text": "Block the card ending 7731"})
n.status # "completed" — the model repliedn.verified # ⚡ was the ACTION confirmedUnderstand the difference — status and verified are not the
same thing. Integrating without knowing this is the most common
mistake: the one rule that matters.
| Ordinary AI pipeline | Davirix | |
|---|---|---|
| Network dropped — did it happen? | ❓ doesn’t know | UNKNOWN, retry forbidden |
| Retry → two SMS? | ⚠ yes | 10,000 parallel → exactly 1 |
| Model said “I sent it” | ✅ believes it | ACKNOWLEDGED ≠ VERIFIED |
| Two years later: “why this decision?” | ❌ no trace | offline manifest, one changed byte is detected |