Reeflex

Reeflex is an open-source AI agent governance gate that decides on every agent action before it reaches a system’s data. It is described as deterministic and uses a four-stage pipeline to intercept an intended action, normalize it into an Action Envelope, apply OPA/Rego policy rules, and then allow, hold, or deny the action. The product is meant to act as a seatbelt for AI acting on connected systems.

Its policy layer evaluates reversibility, blast radius, and externality, and the site says the decision path uses classical logic rather than prompts, models, or other black boxes. The default policy is presented as five readable rules: reversible reads pass, irreversible broad actions in production wait for a human, irreversible systemic actions in production are refused outright, fragmented deletes are caught cumulatively, and everything else is governed by exception. Reeflex also records the verdict and envelope in an append-only audit log, and each decision can stream to a SIEM in standard syslog format (RFC 5424). It supports an observe mode in which actions are logged but not enforced.

The approvals flow is built around a second decision made by a human or a trusted agent. A held action waits for a designated approver in wp-admin or in an MCP client such as Claude Desktop, after which the action runs. The integrations section names a WordPress/WooCommerce plugin, an adapter for Claude Code, an MCP gateway, a self-hosted MCP path, an open Adapter Specification for any backend that can make HTTP calls, and an n8n node called Reeflex Gate. The site also says it ships as a thin adapter per platform.

Reeflex is listed as Apache-2.0 and described as open core. The site also links to GitHub, docs, a quickstart, and a whitepaper PDF.