Decision envelope
The accountable context around one consequential AI decision or agent action.
DEVELOPER QUICKSTART
Record one consequential decision. Verify the evidence.
LoopGrid is SDK/API-first. Start locally, capture one decision path, export the evidence and verify it outside the application.
01 · INSTALL
Choose an SDK.
Python
pip install loopgrid
TypeScript / JavaScript
npm install @cybertechsoft/loopgrid
The current public release supports local and self-hosted evaluation. Point the SDK at your LoopGrid service; deployment-specific security and operational controls are documented on the Security page.
02 · RECORD
Capture the decision path, not just the model call.
from loopgrid import LoopGrid
grid = LoopGrid(
base_url="http://localhost:8000",
api_key="lg_dev_…"
)
grid.record_decision(
decision_type="customer_refund",
privacy_mode="redacted",
agent={"id": "support-agent-04"},
proposed_action={
"tool": "stripe.refunds.create",
"amount": 720
}
)A production workflow can append policy evaluation, human review, external action and observed outcome as signed history rather than rewriting the original decision.
03 · CORE CONCEPTS
The evidence model.
Append-only events
Reviews, corrections, actions, outcomes and replays are added as new history.
Cryptographic proof
SHA-256 commitments, workspace-scoped chain links and signing protect integrity/provenance.
Privacy modes
FULL, REDACTED and PROOF-ONLY let teams choose how much source data LoopGrid retains.
04 · PLATFORM GUIDES
Connect your workflow platform.
Install the published plugin or node, connect your LoopGrid workspace and record evidence around the steps that matter.
VERIFIED DIFY MARKETPLACE
LoopGrid for Dify
Install the verified Dify Marketplace plugin, configure Base URL, Workspace ID and API Key, and use six tools in Workflows or Chatflows.
Dify quickstart →VERIFIED N8N NODE
LoopGrid for n8n
Use LoopGrid directly on n8n Cloud or install n8n-nodes-loopgrid on self-hosted n8n, then capture and verify evidence in the workflow.
LANGFLOW / LFX · PYPI
LoopGrid for Langflow
Install lfx-loopgrid==0.1.0 and add six LoopGrid components to Langflow flows.
LIVE JEV · VERCEL AI GATEWAY
LoopGrid × Jev
Preserve live typed Jev answers, deterministic application policy, delegated authority, sandbox action/outcome and signed verification.
Jev quickstart →STRANDS AGENTS · NATIVE PLUGIN · PYPI
LoopGrid for Strands
Install loopgrid-strands==0.1.0 and attach a native LoopGrid plugin to Strands Agents with request-scoped evidence through invocation_state.
GOOGLE ADK · NATIVE BASEPLUGIN · PYPI
LoopGrid for Google ADK
Install loopgrid-google-adk==0.1.0, attach the native BasePlugin at the ADK App boundary, and keep external execution/outcome evidence explicit.
VERCEL AI SDK 7 · NATIVE TELEMETRY · NPM
LoopGrid for Vercel AI SDK
Install loopgrid-vercel-ai-sdk@0.1.0 and attach decision-scoped native AI SDK telemetry while keeping authority, policy, reviewer identity and observed outcome explicit.
OPENAI AGENTS SDK · NATIVE TRACING + RUNHOOKS · PYPI
LoopGrid for OpenAI Agents SDK
Install loopgrid-openai-agents==0.1.0, bind decision-scoped trace metadata and RunHooks, and keep authority, reviewer identity and observed downstream outcome explicit.
HAYSTACK 3.3 · NATIVE TRACING + AGENT HOOKS · PYPI
LoopGrid for Haystack
Install loopgrid-haystack==0.1.0, enable explicit Haystack tracing, bind decision-scoped Agent hooks, and keep reviewer identity and observed downstream outcome explicit.
CREWAI 1.15.25 · NATIVE EVENT LISTENER · PYPI
LoopGrid for CrewAI
Install loopgrid-crewai==0.1.0, capture CrewAI model and tool events within a decision scope, and preserve explicit review and outcome evidence.
MICROSOFT AGENT FRAMEWORK · NATIVE MIDDLEWARE · PYPI
LoopGrid for Microsoft Agent Framework
Install loopgrid-microsoft-agent==0.1.0 for native Python chat/function middleware, host-supplied review evidence and signed Core verification.
PYDANTIC AI 2.x · NATIVE CAPABILITY · PYPI
LoopGrid for Pydantic AI
Install pydantic-ai-loopgrid==0.1.0 and attach a native capability to capture model and tool execution evidence, enforce policy and human approval, and verify signed outcomes.
CLAUDE AGENT SDK · NATIVE TOOL HOOKS · PYPI
LoopGrid for Claude Agent SDK
Install loopgrid-claude-agent==0.1.0 to enforce exact tool authorization in native Claude SDK hooks, record approval and outcome evidence, and verify signed Core records.
LANGCHAIN · NATIVE AGENTMIDDLEWARE · PYPI
LoopGrid for LangChain
Install loopgrid-langchain==0.1.0 to guard real LangChain create_agent sync/async tool calls with Core policy, review and signed evidence, without mistaking tool results for external receipts.
LANGGRAPH · NATIVE STATEGRAPH + CHECKPOINTS · PYPI
LoopGrid for LangGraph
Install loopgrid-langgraph==0.1.0 for native decision, guarded execution and verification nodes, policy/review gates, checkpoint recovery and signed Core evidence.
MASTRA · NATIVE TOOL GUARDS + WORKFLOWS · NPM
LoopGrid for Mastra
Install loopgrid-mastra@0.1.0 to bind Mastra tool and workflow events to one decision, with Core policy checks, approval provenance and signed verification.
05 · INTEGRATION PATHS
Use the path that fits the system you already run.
| Path | Use |
|---|---|
| Dify | Verified Marketplace plugin with six workflow tools |
| n8n | Verified n8n Cloud node + npm package for self-hosted n8n |
| Langflow | Published PyPI/LFX extension with six evidence components |
| TypeSafe AI Jev | Live Jev decision + application-policy + action/outcome evidence through Vercel AI Gateway |
| Strands Agents | Native Plugin + typed lifecycle hooks + request-scoped invocation evidence |
| Google ADK | Native BasePlugin + per-run binding + explicit tool-result / action-execution boundary |
| Vercel AI SDK | Native decision-scoped Telemetry + explicit human-review/outcome boundary |
| OpenAI Agents SDK | Native tracing + decision-scoped RunHooks + approval-before-execution evidence boundary |
| Haystack | Native Tracer/Span + Agent hooks + ConfirmationHook-safe tool evidence boundary |
| CrewAI | Native CrewAI event listener + tool guardrail-aware capture + explicit review and outcomes |
| Microsoft Agent Framework | Native chat and function middleware with approval and outcome evidence |
| Pydantic AI | Native capability with exact tool authorization, human approval and signed outcome evidence |
| Claude Agent SDK | Native PreToolUse/PostToolUse hooks with exact authorization, human review and signed action/outcome evidence |
| Mastra | Native guarded createTool execution and workflow steps, with decision-scoped Core policy and signed evidence |
| LangGraph | Native StateGraph decision/policy/guard nodes, durable checkpoint recovery, signed approval and replay protection |
| LangChain | Native AgentMiddleware synchronous/asynchronous guarded tool execution, signed policy/review evidence and distinct external receipts |
| Agno | Published Python integration using native hooks for decision evidence |
| REST | Direct evidence API integration |
| Python SDK | Python services and AI backends |
| TypeScript SDK | Node / TypeScript services |
| OpenTelemetry / OTLP | OTLP/HTTP protobuf + JSON trace ingestion, optional gzip |
| MCP proxy | Evidence around tool calls and higher-risk action boundaries |
| GitHub Actions | Verify exported evidence in CI with signer pinning |
06 · OTLP + CI
Use standard telemetry transport. Verify the result independently.
LoopGrid v0.8.1 accepts OTLP/HTTP binary protobuf or JSON on /v1/traces. Many OpenTelemetry SDKs use http/protobuf by default; JSON remains supported for compatibility.
export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://127.0.0.1:8000/v1/traces export OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf export OTEL_EXPORTER_OTLP_TRACES_HEADERS="X-LoopGrid-Key=lg_live_...,X-LoopGrid-Workspace=default"
For CI verification, use the published Marketplace Action against an exported evidence ZIP. Pin an expected public key when signer identity matters.
- uses: actions/checkout@v7
- uses: loopgridio/loopgrid-evidence-verify@v1.0.0
with:
evidence: ./evidence/evidence.zip
trusted-public-key: ./keys/loopgrid-production-public-key.pemVerification establishes integrity/provenance under the expected key for captured records. It does not by itself prove that the underlying decision was correct, that every upstream statement was true, or that capture was complete.
06 · NEXT
Understand the trust boundary before sensitive production use.
Read the public threat model, current readiness matrix and known per-pilot hardening items before treating v0.8 as a production control.