sentinel-agent-sdk

Non-blocking AI agent event logging for Python. Every agent action fires a tamper-evident audit event to Sentinel — without adding latency to your agents.

Install

pip install sentinel-agent-sdk

Requires Python 3.9+. Zero dependencies (stdlib only).

Quick start

import sentinel

# Initialize once at startup
sentinel.init(api_key="sk_live_...")

# Option 1 — decorator (auto-logs every call)
@sentinel.wrap("payment-agent", "invoke")
def run_agent(prompt: str) -> str:
    ...  # your agent logic
    return result

# Option 2 — log events manually
sentinel.log_event(
    "payment-agent",
    "tool_call",
    duration_ms=120,
    input_data={"tool": "stripe_charge", "amount": 2999},
    output_data={"status": "ok"},
)

Events fire in daemon threads — your agent never waits for Sentinel.

Configuration

API

sentinel.init(api_key, api_url=None) → SentinelClient

Initialize the default client. Call once at startup.

sentinel.log_event(agent_id, action_type, *, input_data=None, output_data=None, duration_ms=None, framework=None, session_id=None)

Log a single event. Returns immediately (fires in background thread).

sentinel.wrap(agent_id_or_fn, action_type="invoke", *, framework=None)

Wrap a callable to auto-log every invocation. Supports two forms:

# Decorator with explicit IDs
@sentinel.wrap("my-agent", "process")
def my_fn(...): ...

# Direct wrap — agent_id inferred from function name
wrapped = sentinel.wrap(my_fn)

SentinelClient

For multi-tenant apps or multiple API keys:

from sentinel import SentinelClient

client = SentinelClient(api_key="sk_live_...")
client.log_event("agent-a", "invoke")

Field stripping

The following fields are automatically stripped from input_data and output_data before sending to protect sensitive data: prompt, completion, content, messages, text.

License

MIT