Skip to content

Getting Started

Start with deterministic tool-call supervision using the 0.9.0 API (unchanged from 0.8.0). These examples run without a model, framework installation, or API key.

Installation

Requires Python 3.10 or newer. Install the v0.9.0 package from PyPI:

pip install agentic-sidecar==0.9.0
# Optional live Judge SDKs:
pip install "agentic-sidecar[openai,anthropic]==0.9.0"

For development from a checkout, run uv sync --extra dev --extra langgraph --extra docs.

Observe a proposed action

from agentic_sidecar import DecisionContext, Sidecar
from agentic_sidecar.gate.policy import PolicyAdvisor, PolicyRule

sidecar = Sidecar(
    on_sidecar_failure="fail_closed",
    policy=PolicyAdvisor([PolicyRule(tool="delete_*", effect="deny")]),
)
decision = sidecar.evaluate(DecisionContext(tool_name="delete_customer"))
print(decision.status)  # BLOCK

evaluate() computes and records the verdict. Execution is handled by the caller or an adapter. Observe mode lets wrapped tools run regardless of verdict.

Enforce an intent constraint

from agentic_sidecar import Sidecar, SidecarBlockedError
from agentic_sidecar.adapters.langgraph import attach
from agentic_sidecar.intent import ConstraintBinding, IntentEnvelope, IntentGuardian, Requester

def issue_refund(order_id: str, amount: float) -> str:
    """Issue a refund for an order."""
    return f"refunded {amount} for {order_id}"

envelope = IntentEnvelope(
    goal="Refund order A100 up to $500",
    requested_by=Requester(id="support_lead"),
    constraints={"maximum_refund": 500},
)
guardian = IntentGuardian(envelope, [ConstraintBinding(
    constraint="maximum_refund", tool="issue_refund", arg_name="amount", op="lte",
)])
sidecar = Sidecar(
    on_sidecar_failure="fail_closed", mode="govern",
    intent=guardian, roles=["policy", "risk", "intent_guardian"],
)
(refund,) = attach(sidecar, [issue_refund])
print(refund("A100", 120))  # refunded 120 for A100
try:
    refund("A100", 850)
except SidecarBlockedError as exc:
    print(exc.decision.status)  # BLOCK; refund function was not called

The wrapper works with plain callables. To build an actual LangGraph agent, install agentic-sidecar[langgraph] and run the scripted graph examples in Examples. They include a model fixture and graph invocation.

Govern permits ALLOW/WARN, stops BLOCK, resolves PAUSE/ESCALATE through a registered handler, and suspends unresolved outcomes plus CHALLENGE/REPLAN. See Governance and supported scope.

Next steps

  • Configuration: roles, policy/risk rules, intent, budget.
  • Features: evaluator and narration usage.
  • Architecture: boundary responsibilities and decision flow.
  • CLI reference: local status and simulated demo commands.
  • Roadmap: live status in 0.7 and additional adapters in 0.8.

See v0.9 authority, privacy and migration for the current lifecycle APIs.