As the Model Context Protocol rapidly gains adoption across the AI tool ecosystem, one question looms: who watches the tool calls? Enter Bulwarkai, a Unix daemon that sits transparently between MCP clients and downstream servers, inspecting and enforcing security policy on every interaction that passes through it.
The firewall features a six-stage inspection pipeline, SHA-256 tool schema pinning to prevent unauthorized tool modifications, a sliding-window rate limiter, circuit breaker pattern for graceful degradation, and full audit telemetry in JSONL format. Twenty-two headers, 30 source files, and 19 test files shipped in the initial checkin—a complete, functional MVP delivered in a single burst.
Built in C11 with CMake, leveraging cJSON and libuv, the daemon handles stdio transport and full MCP protocol negotiation including capability handshaking. It’s the kind of project that makes security engineers nod approvingly and everyone else ask, “Wait, you wrote a firewall in C… on purpose?”
In an era of rapid MCP adoption where tool calls can read files, execute code, and access sensitive APIs, Bulwarkai represents a critical missing layer—a policy enforcement point that treats AI tool interactions with the same rigor applied to network traffic.
“It was easier to VIBE this than confirm PatchCleaner.exe from the Internet was safe.”
— PatchCleanerAI README
A comprehensive Windows supply chain security assessment tool reached MVP status this month, capable of performing 80 security checks across 19 categories on Windows 11 machines. From firmware validation to executable integrity to OEM pre-install analysis, the tool produces severity-scored findings in both JSON and HTML reports. Golden image fingerprinting enables baseline comparison for fleet deployments, answering the question every security team asks: has this machine been tampered with before it reached our desk?
Somewhere between a love letter to 1991 and a feat of modern Python engineering, Dungeons of Dreagoth emerged this month as a fully playable terminal dungeon crawler sporting procedurally generated levels, D&D-style mechanics, and an AI Dungeon Master powered by Claude.
The game features an 80×40 grid with 25 rooms per level, fog of war with darkvision support, four character classes, four races, and an equipment system spanning 84 items including procedurally generated magic weapons and armor. Nine commits in March refined everything from magic item generation to combat balance to the quest system.
The AI narrator provides atmospheric descriptions of rooms, encounters, and loot—but falls back gracefully to handcrafted templates when offline, because even adventurers sometimes lose their connection to the ethereal plane.
Built on the Textual and Rich frameworks with numpy for dungeon grid generation, the game also features SQLite-backed AI response caching to keep the Dungeon Master’s token bills under control during extended crawling sessions. A full 345-test suite ensures that no goblin spawns where it shouldn’t.
A new Python tool for the discerning podcast listener arrived this month. Podcastorum fetches podcasts via RSS, transcribes them locally using GPU-accelerated Whisper—no audio ever leaves the machine—and then synthesizes editorial-style analysis through a cooperative and adversarial multi-LLM pipeline via LiteLLM.
The architecture is privacy-first by design. March saw the initial framework land, followed by improved Whisper failure handling and the podcast RSS fetcher coming online across four focused commits.
The multi-AI comparison tool—which sends identical prompts to every major LLM provider and orchestrates a comprehensive evaluation pipeline—expanded its roster in March with Mistral integration. A score parsing bug was squashed along the way, ensuring that the adversarial debate rounds produce consistent numerical scores across all providers.
The tool’s pipeline now covers pointwise scoring, pairwise evaluation, adversarial debate, collaborative consensus, and Bradley-Terry ranking across six providers, including support for local GGUF models via llama-cpp-python. PDF report generation delivers polished results suitable for stakeholder review.
With Mistral joining OpenAI, Claude, Gemini, Grok, and local models, the tool now offers the broadest provider coverage of any open-source LLM evaluation framework—and the adversarial layer means no single provider gets to grade its own homework.
Rather than download an unknown .exe from the internet to clean orphaned Windows Installer files, Ron did what any reasonable person would do: built his own replacement from scratch. PatchCleanerAI scans C:\Windows\Installer, identifies orphaned .msi and .msp files through registry and COM API queries, and optionally invokes Claude for files that resist classification.
The tool handles admin privilege elevation, offers safe archiving before deletion, and logs every decision. Three commits in March took it from initial concept to documented, deployable utility—a small project with an outsized lesson about supply chain trust.
The multi-agent open source intelligence tool—which queries 39+ public reconnaissance sources and feeds results through cooperative-adversarial AI analysis—continued refinement into March. The system’s domain-pointed intelligence pipeline now covers infrastructure, personnel, products, financials, security posture, and corporate structure, all assembled from publicly available data with multi-provider AI cross-validation.
The Windows supply chain security assessment tool, which shipped its 80-check MVP in February, continued to mature in March with additional golden image fingerprinting refinements and deployment improvements, solidifying its position as a practical tool for verifying laptop integrity before network onboarding.
··· “Frustrating adversaries since the dial-up era” · GitHub: rondilley · 42 Repositories and Counting ···
The Jetson Orin Nano development board received Python scripts for its Yahboom MINI CUBE NANO case this month, featuring I2C control of 14 WS2812B RGB LEDs and an SSD1306 OLED display. LED effects include breathing, marquee, and rainbow patterns, while the OLED shows live CPU, RAM, disk, and network statistics. A systemd service ensures the display runs on boot—because an AI development board without rainbow LEDs is just a board with an identity crisis.
An M3U/M3U8 playlist resolver that matches broken playlist entries to actual music files on disk saw active development right up to press time. The tool handles the eternal mismatch between “Artist - Title.mp3” and “Artist/Album/01 Title.m4a” through a five-level fallback: exact path, filename match, stem match, ID3 metadata tag reading, and fuzzy matching with false-positive protections including stopword filtering, coverage gates, and artist mismatch caps. It supports over 15 audio formats.
Thirty-plus program suites from 1991–1995 were committed to git this month—Ron’s original DOS-era utilities including RADLab for BBS archive testing, AirTouch Credit System, Head-2-Head multiplayer gaming, and TimeCard employee tracking. Over 300 source files in BASIC, C, and C++ across multiple distribution disks now live in version control. A digital time capsule from the Borland compiler era, preserved for posterity.
A command-line sentinel for Global Privacy Control. Checks .well-known/gpc.json endpoints, analyzes cookie behavior with and without Sec-GPC headers (usprivacy, OptanonConsent), and detects response differences. A quiet but important tool for CCPA/CPRA compliance verification.
Both projects saw activity in March, exploring AI-assisted data visualization and mathematical pattern recognition. Prime Plot focuses on novel visual representations of prime number distributions, while StarPattern applies pattern detection to astronomical and geometric data sets.
The OSINT AI project continued to mature through February and into early March, building an AI-augmented open source intelligence gathering and analysis pipeline for security researchers and analysts.
Ron’s omarchy-retropc-theme remains the most-starred personal project at 66 stars, with the arc_raiders and himalaya themes also drawing community attention. The retro PC aesthetic continues to resonate with the Omarchy desktop community.