Verifiable decisions for LangGraph agents.

Install the released Python integration.

python -m pip install loopgrid-langgraph==0.1.0

The package provides native StateGraph decision, guarded-execution and verification nodes via LangGraphBridge. It requires Python 3.10+ and declares langgraph>=1.2.12,<2 and httpx>=0.27,<1. The persistent SQLite test optionally installs langgraph-checkpoint-sqlite. The package is community-built, not an officially endorsed LangGraph extension.

Bind one proposed action to Core, then guard the native node.

from typing import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.types import Command
from loopgrid_langgraph import CoreClient, LangGraphBridge
import os

class State(TypedDict, total=False):
    amount: int
    operation_id: str
    loopgrid: dict
    loopgrid_tool_result: dict
    loopgrid_verification: dict

core = CoreClient(
    base_url="http://127.0.0.1:8000",  # loopback only; use HTTPS remotely
    workspace_id="YOUR_TEST_WORKSPACE",
    service_key=os.environ["LOOPGRID_SERVICE_KEY"],
)
bridge = LangGraphBridge(core, agent_id="refund-agent")
arguments = lambda s: {"amount": s["amount"], "currency": "USD"}

propose = bridge.make_decision_node(
    tool="refund", policy_id="refund-policy", arguments=arguments,
    authority=lambda s: {"limit_usd": 1500, "scope": ["refund:create"]},
    model=lambda s: {"name": "host-supplied-model-identity"},
    context=lambda s: {"prompt_version": "refund-v1"},
    operation_id=lambda s: s["operation_id"],
)
# Only synthetic local data. No payment or refund API is called.
execute = bridge.make_guarded_node(
    tool="refund", arguments=arguments,
    callback=lambda args: {"sandbox": True, "amount": args["amount"]},
)

builder = StateGraph(State)
builder.add_node("propose", propose)
builder.add_node("execute", execute)
builder.add_node("verify", bridge.make_verify_node())
builder.add_edge(START, "propose")
builder.add_edge("propose", "execute")
builder.add_edge("execute", "verify")
builder.add_edge("verify", END)
graph = builder.compile(checkpointer=InMemorySaver())
cfg = {"configurable": {"thread_id": "refund-thread-unique-01"}}
result = graph.invoke({"amount": 75, "operation_id": "refund-op-unique-01"}, cfg)

# When policy requires approval, result includes an interrupt.
# ONLY the authenticated host requests approval through Core's review API.
# Graph resume text is never a human approval credential.
# After authorized review is recorded against the SAME decision:
# result = graph.invoke(Command(resume="continue"), cfg)

This is a synthetic refund example, not a payment application. Create a disposable workspace, configure a real Core policy and supply a scoped service credential. InMemorySaver is for tests; use a durable supported checkpointer and a host-provided dispatch lease in production.

Evidence follows the consequential action, not the graph's narration.

make_decision_node(), make_guarded_node() and make_verify_node() return native graph nodes. LoopGridGraphCallback is for non-authorizing interrupt/resume observations only.

Checkpoint restoration must not repeat a consequential action.

The integration does not guarantee globally exactly-once business effects. In multi-worker deployments, use a host-controlled atomic distributed lease or outbox. The included SQLite lease is a single-host reference, not a distributed lock.

Keep checkpoints, credentials and commitments under application control.

The loopgrid graph scope avoids duplicating the raw proposed tool-argument map. Other host state, model outputs, callback results and checkpoint stores may still contain sensitive data; deploy retention, encryption and access controls accordingly. Core redacted fields use typed SHA-256 commitments, which are checked rather than ignored on replay.

Tested against native LangGraph and LoopGrid Core.

These tests use controlled sandbox actions and do not constitute certification for real-money workflows. Signed workspace verification is not itself a claim that an independently exported bundle was verified offline in the same run. See the validation record ↗ and integration contract ↗ for boundaries.

LangGraph orchestrates. Your host authorizes. LoopGrid records evidence.

Do not equate agent text, graph resume values, successful callbacks or signed histories with a verified transfer or refund. Keep authentication, real external action execution, downstream reconciliation, signer trust and checkpoint privacy under independent application control.

Source: GitHub ↗ · Security notes ↗ · PyPI ↗.