LoopGrid keeps a provider-neutral evidence model while exposing several capture paths. Choose the least invasive integration that still captures the claim you need to verify.
Connect through published integrations for Dify, n8n, Langflow, Strands Agents, Google ADK, Vercel AI SDK, OpenAI Agents SDK, Haystack, CrewAI, Microsoft Agent Framework, Pydantic AI, Claude Agent SDK, Mastra, LangGraph, LangChain and Agno, or the live-validated Jev community integration. Each path uses the same LoopGrid evidence contract.
Verified by Dify. Six LoopGrid tools for Workflows and Chatflows record decisions, external actions, outcomes and reviews, then verify the signed evidence trail.
Verified by n8n and live on n8n Cloud. Record decisions, external actions, outcomes and human reviews in Cloud or self-hosted workflows, then verify the signed evidence trail.
Six published Langflow components for recording decisions, actions, outcomes and reviews, then verifying the signed evidence trail.
Live-validated community integration for preserving typed Jev decisions, deterministic application policy, delegated authority, sandbox action/outcome and cryptographic verification.
Native Strands plugin using typed lifecycle hooks to capture decision, model, tool and observed-outcome evidence without taking over agent execution.
Native Google ADK plugin for recording run, model and tool lifecycle evidence while keeping authoritative external-action and outcome claims application-owned.
Decision-scoped native Telemetry integration for model and tool lifecycle evidence, with authority, policy, reviewer identity and downstream outcome kept explicit and application-owned.
Decision-scoped evidence and Core policy gating inside native Mastra tool execution, with host-authenticated human review and independently observed external outcomes.
Native create_agent tool guards with signed decision and policy evidence, host-authenticated review, conservative replay protection, and separate tool results and external outcomes.
Native graph decision, policy/review and guarded execution nodes with persistent checkpoint recovery, conservative replay protection and signed Core evidence.
Native tracing for model evidence plus decision-scoped RunHooks at the actual local tool boundary, with human approval, authority, policy and downstream outcome kept explicit.
Native Haystack integration combining haystack.agent.step.llm tracing with Agent tool hooks, ConfirmationHook-safe ordering and explicit human-review/outcome boundaries.
Capture native CrewAI model completions and actual tool executions as signed decision evidence, with explicit policy, human review and observed outcome.
Published Python integration using Microsoft’s native chat and function middleware. Capture model and tool events, application policy and host-supplied human review, with observed outcomes recorded in the existing LoopGrid Core.
Native Pydantic AI capability captures model and tool lifecycle evidence, enforces exact action authorization and supports deferred human approvals. Verify signed decisions and observed outcomes through LoopGrid Core.
Published Python integration for Claude's native tool lifecycle. Apply decision-scoped policy and human-approval gates before tool execution, capture the actual result or failure, and verify signed evidence through LoopGrid Core.
The published Agno integration connects framework hooks to LoopGrid decision evidence. See the package documentation for supported events and the current validation scope.
These integrations record and verify evidence; they do not execute refunds, transfers or other real business actions. The Jev, Strands, Google ADK, Vercel AI SDK, OpenAI Agents SDK, Haystack, CrewAI, Microsoft and Claude Agent SDK examples use sandboxed actions with no real money movement; the Claude SDK's real model-driven tool test remains separately documented in its validation record. Mastra was additionally tested with real provider-backed agent generate/stream calls and synthetic tool actions. LangGraph passed real StateGraph interrupt/resume, persistent SQLite checkpoint recovery and local Core signed-policy/review E2E, also using sandbox actions. LangChain passed native sync/async AgentMiddleware execution, replay protection and combined real Core signed E2E with synthetic tools.
A simple application event and a system-of-record action do not need the same integration. Higher-risk claims should capture closer to the consequential action boundary.
Record decisions and events directly from the application using REST, Python or TypeScript.
Best for:Explicit decision lifecycle instrumentation.Accept OTLP/HTTP protobuf or JSON on /v1/traces, with optional gzip, then map compatible telemetry into the same evidence path.
Capture the consequential tool request and downstream result when the evidence depends on an external state change.
Best for:Agent actions where “what actually happened” matters.Export a portable evidence bundle and verify captured integrity/provenance offline or gate CI with the published GitHub Action.
Best for:Customer, security, audit, release and investigation workflows.Dify, n8n, Langflow, Jev, Strands, Google ADK, Vercel AI SDK, OpenAI Agents SDK, Haystack, CrewAI, Microsoft Agent Framework, Pydantic AI, Claude Agent SDK, Mastra, LangGraph, LangChain and Agno use the same underlying decision/event evidence contract as the core SDK/API and open-protocol paths. Choose the integration that fits your workflow.
The public loopgrid-evidence-verify@v1.0.0 Action runs LoopGrid's offline verifier in CI and can pin an expected signer key. It verifies captured integrity/provenance; it does not prove decision correctness or capture completeness.
Beyond the published integrations listed above, additional ecosystem work may explore expanded MCP discovery, Pipedream, Make, Zapier and other provider-specific adapters as demand develops.
If the evidence claim is that an external action really occurred, capture the request and authoritative response at a boundary the agent cannot rewrite. Correlation, retries, idempotency and failure handling should remain explicit.
Use a durable decision/workflow ID across agent, approval and external systems.
Record the consequential request and observed downstream result at the right boundary.
Export evidence and check integrity/provenance independently using the expected signer identity.
Start locally, instrument one workflow, export the evidence, then decide whether a deeper adapter is justified.