humanlatch 0.1.0

Python SDK for proposing AI-driven actions to a HumanLatch control plane.

HumanLatch Python SDK

Python SDK for sending AI-driven capability proposals to a HumanLatch control plane.

Use this SDK when your agent, service, robot controller, or automation needs to ask:

• Can I run this action automatically?
• Do I need a human approval first?
• Is this blocked by policy?

The SDK works against either:

• self-hosted HumanLatch CE
• a remotely hosted HumanLatch API

Install

pip install humanlatch

Create a Client

from humanlatch import HumanLatchClient

client = HumanLatchClient(
    base_url="https://your-humanlatch.example.com",
    workspace_id="your-workspace-id",
    api_key="hl_your_key_here",
)

Use token for user-authenticated flows or api_key for agent/service integrations.

Core Flow

from humanlatch import HumanLatchClient

client = HumanLatchClient(
    base_url="https://your-humanlatch.example.com",
    workspace_id="your-workspace-id",
    api_key="hl_your_key_here",
)

result = client.propose_action(
    action_type="support.refund.issue",
    target="order-10428",
    summary="Issue a refund above the autonomous threshold",
    payload={
        "amount": 850,
        "currency": "USD",
        "customer_id": "cust_2041",
    },
    context={
        "environment": "production",
        "requested_by": "agent:support-bot",
        "domain": "chatbot",
    },
)

if client.is_approved(result):
    issue_refund()
elif client.requires_approval(result):
    final = client.wait_for_decision(result["id"], interval_seconds=2.0, timeout_seconds=30.0)
    if client.is_approved(final):
        issue_refund()
else:
    raise RuntimeError(f"Blocked by HumanLatch policy: {result['status']}")

Additional Methods

client.list_actions(status="pending_approval", limit=50)
client.get_action("action-id")
client.approve_action("action-id", note="Approved by supervisor")
client.reject_action("action-id", note="Unsafe during shift handoff")
client.cancel_action("action-id")

Integration Model

Before your system performs a sensitive capability, it asks HumanLatch for a decision:

• approved_auto: execute now
• pending_approval: wait for human approval
• blocked: do not execute

That same control plane works for:

• infrastructure agents
• support bots
• warehouse robots
• manufacturing lines
• internal copilots