sentinel-llm 0.1.0

Drop-in observability and trust layer for AI pipelines and agents.

sentinel-llm

Drop-in observability and trust layer for AI pipelines and agents. Part of Sentinel. Installs as sentinel-llm, imports as sentinel.

The SDK has zero runtime dependencies (standard library only) so it's safe to import into any process. It intercepts LLM calls, tool calls, and retrieval steps and ships them - batched, PII-redacted, on a background thread - to a Sentinel collector.

Install

pip install sentinel-llm
# optional provider extras
pip install "sentinel-llm[openai]"      # or [anthropic], [otel]

Use

import sentinel
from openai import OpenAI

sentinel.init(service_name="my-agent")     # defaults to http://localhost:8000
client = sentinel.wrap(OpenAI())           # the only change to your code

client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)

Instrument your own code:

@sentinel.trace(kind="tool")
def search(query: str) -> list[str]:
    ...

with sentinel.span("agent-run", kind="agent"):
    search("vector databases")             # nested under the agent span

Works with sync and async clients, streaming responses, and Anthropic (sentinel.wrap(Anthropic())). For other frameworks, use @sentinel.trace / sentinel.span() directly.

Configuration

All settings have env-var equivalents (SENTINEL_*); see the configuration guide.

Option	Default	Purpose
endpoint	http://localhost:8000	collector URL
service_name	default	logical service name
capture_content	true	record prompt/response payloads
redact_pii	true	redact PII before export
sample_rate	1.0	head-based sampling
enabled	true	set false to make the SDK a no-op

License

Apache 2.0.