MCP Local Auditor & System Architect V2

The server automatically detects whether an incoming submission is code or a high-level specification and routes it to one of two roles. In Production Code Auditor mode, it checks the input against ten defensive engineering rules covering logic and edge conditions such as null checks and division by zero, resource leaks and performance issues such as unclosed context managers, and defensive boundaries such as type hints and bounded loop guards. It returns a deterministic APPROVED or REJECTED verdict with actionable feedback. In System Architect & Strategist mode, triggered when the input looks like a specification rather than code, it produces component topologies, directory structures, risk modeling for concurrency and failure points, and a phased execution plan.

Before anything reaches the language model, a fast AST-based sentinel parses the submission and blocks high-risk calls such as eval, exec, os.system and subprocess, and rejects malformed Python outright. Every audit is recorded twice: as structured rows in a local SQLite database for metrics and history, and as atomic JSON logs meant to double as a fine-tuning dataset. A separate watcher script can run alongside the server and post formatted status updates to a Discord channel as new audit files appear, so a team can follow review activity without opening a dashboard. The server also queries LM Studio's active model endpoint directly, so it picks up whichever model is currently loaded rather than requiring a hardcoded model name.

The project ships a ready-made "Agent Loop Protocol" snippet that developers can paste into their coding assistant's configuration, instructing the assistant to call the audit tool after every code change, read the rejection feedback, fix the reported issues, and resubmit, capped at five attempts.

It is built for individual developers and small teams who want an automated code and architecture review step in their MCP-based coding workflow without paying for cloud inference or exposing proprietary code to a third-party API. It requires Python 3.10 or later, a running LM Studio instance with an instruction-tuned model loaded, and an MCP-compatible client. The project is distributed under the MIT license and is free to use; the maintainer accepts optional donations through Ko-fi and GitHub Sponsors.