LoopGrid

LoopGrid is decision evidence infrastructure for AI agents, published by cybertechsoft. It captures the agent, delegated authority, model and policy context, human approval, external action and observed outcome of a consequential decision, then seals them as signed, tamper-evident records that can be verified independently. The current release is v0.8.1, positioned for controlled design-partner technical evaluation rather than production GA.

What It Is

LoopGrid is an evidence layer that sits alongside agent runtimes, observability and governance tools rather than replacing them. Teams record a decision through REST, Python or TypeScript SDKs, OpenTelemetry/OTLP or MCP. Each evidence event is committed to a workspace-isolated SHA-256 hash chain and signed (Ed25519 locally, with an AWS KMS path). Reviews, actions, outcomes and replays are appended as new signed events instead of rewriting the original decision.

Capture to proof workflow

The lifecycle runs capture, seal, investigate and prove. Investigators can reconstruct what happened, inspect approvals and policy versions, and run counterfactual replay without mutating production history. Exports are portable evidence bundles containing a manifest, decision and event files, chain witnesses, a public key and an HTML report. A standalone verifier (loopgrid-verify) checks hashes, signatures and chain continuity offline, with signer pinning, and a GitHub Action can verify exported bundles in CI. Verification establishes integrity and provenance; it does not prove a decision was correct or confer legal compliance.

Privacy modes

FULL, REDACTED and PROOF-ONLY modes control how much source data is retained. Raw payloads can sit in a separate AES-256-GCM disclosure vault, be replaced by field commitments, or stay customer-side with only the commitment kept. Encrypted disclosures can be erased without rewriting the signed chain.

Integrations and setup

Documented integrations include a verified Dify Marketplace plugin, an n8n node, Langflow, Strands Agents, Google ADK, Vercel AI SDK, OpenAI Agents SDK, Haystack, CrewAI, Microsoft Agent Framework, Pydantic AI, Claude Agent SDK, LangChain, LangGraph, Mastra and Agno. A no-Docker quickstart script creates a synthetic refund decision, exports a bundle and verifies it; a PostgreSQL-backed Docker path exists for design-partner deployments.

Current Status

The README states the release is approved for controlled design-partner evaluation and is not Production GA. It does not claim a production SLA, legal compliance, validated multi-instance scaling, or live hardware-backed KMS signing unless exercised in a given deployment. The source is Apache 2.0 licensed.