Mneme

Adds a verification and security layer between AI coding assistants and your local codebase. Exposes MCP tools that screen prompts for injection attempts, blind secrets before context reaches the model, gate destructive shell commands with co-sign requirements, and index git history for cross-session memory. The rail commands (ingress/egress) let you policy-check what goes to the LLM and what comes back before it touches your tree. Also includes a repo health grader that scores dependencies, complexity, and bus factor with reproducible metrics. Built for air-gap scenarios where you need signed, offline-verifiable proof of what crossed the wire. Install locally, point your agent at the stdio transport, and every action gets metered and logged.