sentinos 0.1.8

Sentinos Python SDK (control-plane client + runtime governance adapters)

sentinos (Python SDK)

Sentinos is the control plane for AI agents: runtime governance, deterministic policy outcomes, and trace-backed forensics.

This package is the ergonomic Python wrapper over sentinos-sdk-core, exposing operator-first clients for:

• Kernel (execution boundary, autonomy sessions, escalations, traces)
• Arbiter (policy lifecycle + deterministic outcomes)
• Chronos (context snapshots and provenance)
• Alerts, incidents, marketplace, and supporting workflows

Install

pip install sentinos

Optional extras:

pip install "sentinos[providers]"  # openai + anthropic + boto3 (bedrock)
pip install "sentinos[otel]"       # OpenTelemetry helpers
pip install "sentinos[langchain]"  # LangChain integration helpers
pip install "sentinos[grpc]"       # grpcio + protobuf (native gRPC protocol smoke/integration)

Configure

The SDK supports a “single URL” setup by default:

export SENTINOS_BASE_URL="https://api.sentinos.ai"
export SENTINOS_ORG_ID="<org-id>"
export SENTINOS_ACCESS_TOKEN="<access-token>"

Notes:

• SENTINOS_ORG_ID is preferred; SENTINOS_TENANT_ID remains supported as an alias.
• If you run services on separate hosts, set SENTINOS_KERNEL_URL, SENTINOS_ARBITER_URL, SENTINOS_CHRONOS_URL.

Quickstart

Environment-driven:

from sentinos import SentinosClient

client = SentinosClient.from_env()
print(client.kernel.get_runtime_metrics())
print(client.arbiter.governance_dashboard())

Explicit constructor (use org_id; tenant_id is an alias):

from sentinos import SentinosClient
from sentinos.auth.jwt import JWTAuth

client = SentinosClient.simple(
    base_url="https://api.sentinos.ai",
    org_id="acme",
    auth=JWTAuth(lambda: "<access-token>"),
)
print(client.kernel.get_runtime_metrics())

Workforce Auth (Enterprise)

Enterprise workforce token exchange auth:

from sentinos import SentinosClient, WorkforceAssertion, WorkforceTokenProvider
from sentinos.auth.jwt import JWTAuth

workforce_provider = WorkforceTokenProvider.from_env(
    assertion_provider=lambda: WorkforceAssertion(
        external_subject="employee-123",
        email="[e-mail]",
        groups=["AI_USERS"],
    ),
    idp_issuer="https://login.microsoftonline.com/tenant/v2.0",
)

client = SentinosClient(
    org_id="enterprise-org",
    base_url="https://api.sentinos.ai",
    auth=JWTAuth(workforce_provider),
)

Workforce token CLI bootstrap (helpful for enterprise workstation rollout and diagnostics):

sentinos-workforce-auth exchange \
  --controlplane-url "https://app.sentinoshq.com" \
  --org-id "<org-id>" \
  --idp-issuer "https://login.microsoftonline.com/<tenant>/v2.0" \
  --external-subject "<employee-sub>" \
  --assertion-token "<signed-idp-jwt>" \
  --audience "sentinos-workforce"

LLM / Agent Runtime Integration

Use LLMGuard when your application executes provider calls directly (OpenAI, Anthropic, LangChain tools, custom APIs) and you still want Sentinos policy decisions + decision traces for every interaction.

from sentinos import LLMGuard, SentinosClient

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-123")

result = guard.run(
    provider="openai",
    operation="chat.completions",
    model="gpt-4o-mini",
    request={"messages": [{"role": "user", "content": "Summarize this incident"}]},
    invoke=lambda: {"id": "resp-1", "model": "gpt-4o-mini"},
)
print(result.trace.trace_id, result.trace.decision)

For Open Responses-compatible providers/endpoints:

from sentinos import LLMGuard, SentinosClient, create_openresponses_adapter

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-openresponses-1")
adapter = create_openresponses_adapter(guard=guard, client=openai_client)
result = adapter.create(
    model="gpt-4.1-mini",
    input=[{"type": "message", "role": "user", "content": "summarize recent incidents"}],
)
print(result.trace.trace_id, result.trace.decision, result.response.status)

Drop-in adapter classes are also available when you already have provider client objects:

from sentinos import LLMGuard, OpenAIChatCompletionsAdapter, SentinosClient

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-123")
adapter = OpenAIChatCompletionsAdapter.from_client(guard=guard, client=openai_client)
result = adapter.create(model="gpt-4o-mini", messages=[{"role": "user", "content": "hello"}])
print(result.trace.trace_id, result.trace.decision)

Optional extras:

pip install "sentinos[providers]"   # openai + anthropic + boto3 (bedrock)
pip install "sentinos[bedrock]"     # boto3 only
pip install "sentinos[grpc]"        # grpcio + protobuf (native gRPC protocol smoke/integration)
pip install "sentinos[langchain]"   # langchain runtime integrations

Native Kernel gRPC example (non-Go interoperability):

export SENTINOS_GRPC_TARGET="<kernel-grpc-endpoint>"
export SENTINOS_ACCESS_TOKEN="<jwt-access-token>"
export SENTINOS_ORG_ID="<org-id>"
python examples/protocols/grpc_execute_smoke.py

Factory helpers for low-friction org onboarding:

from sentinos import LLMGuard, SentinosClient, create_openai_chat_adapter

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-123")
adapter = create_openai_chat_adapter(guard=guard, client=openai_client)
result = adapter.create(model="gpt-4o-mini", messages=[{"role": "user", "content": "hello"}])
print(result.trace.trace_id, result.trace.decision)

Helper for OpenAI-style chat.completions.create signatures:

from sentinos import LLMGuard, SentinosClient, guard_openai_chat_completion

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-123")

def fake_create(*, model, messages, temperature=0.2):
    return {"id": "chat-1", "model": model, "messages": messages}

result = guard_openai_chat_completion(
    guard=guard,
    create=fake_create,
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "hello"}],
)
print(result.response["id"])

OpenRouter factories (provider identity: openrouter):

from sentinos import LLMGuard, SentinosClient, create_openrouter_chat_adapter

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-openrouter-1")
adapter = create_openrouter_chat_adapter(
    guard=guard,
    api_key="<OPENROUTER_API_KEY>",
    http_referer="https://console.example.com",
    x_title="Sentinos Console",
)
result = adapter.create(model="openai/gpt-4.1-mini", messages=[{"role": "user", "content": "hello"}])
print(result.trace.trace_id, result.trace.decision)

Amazon Bedrock native Converse adapters (provider identity: bedrock):

from sentinos import LLMGuard, SentinosClient, create_bedrock_converse_adapter

client = SentinosClient.from_env(org_id="acme")
guard = LLMGuard(kernel=client.kernel, agent_id="assistant-1", session_id="sess-bedrock-1")
adapter = create_bedrock_converse_adapter(guard=guard, region_name="us-east-1")
result = adapter.converse(
    model_id="anthropic.claude-3-5-sonnet-20240620-v1:0",
    messages=[{"role": "user", "content": [{"text": "summarize incidents"}]}],
)
print(result.trace.trace_id, result.trace.decision)

Reference

• Docs: https://docs.sentinoshq.com/sdk/

Trace Forensics and Retention Example

from sentinos import SentinosClient

client = SentinosClient.from_env(org_id="acme")
trace_id = "11111111-1111-1111-1111-111111111111"

ledger = client.traces.ledger_verify(trace_id)
replay = client.traces.replay_trace(trace_id, request={"include_explain": True})
retention = client.traces.get_retention_policy()
dry_run = client.traces.enforce_retention(request={"dry_run": True})
distributed = client.traces.distributed_trace_summaries(limit=25)

print(ledger.verified, replay.drift_detected, retention.trace_days, dry_run.traces_affected, len(distributed))