I build tools for one habit I have kept since the first chat models: check what the model would otherwise just assert. Some of these tools run AI on hardware you own, with no cloud in the loop. Some are Claude AI skills that make Claude test its own claims before it hands them to you. One is a game. Most are open source. TigerOCR is a hosted service billed by the GPU second, and GlassVGL and LightEDA sell paid packs next to their free parts. Every one is listed in this directory with its license and how it is paid for.

This page is the map. Each section links to a longer guide.

On-board AI: OpenCarMind

OpenCarMind is an open-source OBD2 logger for the ESP32 that runs its neural network on the chip. It reads your car through a Bluetooth LE OBD-II adapter, classifies the last 5 seconds of driving, and estimates brake pressure, flywheel torque and fuel rate, three values the OBD port does not report. It needs no phone, no account and no internet connection. One of its two models keeps learning your car while you drive.

That is what on-board AI means here: the model, the data and the learning all stay on a board in your car. The guide to on-board AI in OpenCarMind explains why that matters, and the EdgeOS article covers how the ESP-Mosaico build keeps every model on time. The full manual is in the OpenCarMind documentation.

Claude AI skills: Skill Vision, Research Hound and Agent Oracle

Three Claude AI skills for Claude Code, published as one suite. Each installs with one git clone into your skills folder, and each is scanned by NVIDIA SkillSpector on every push.

  • Skill Vision is a Claude Agent Skill validator. It tells you whether your other skills will load, upload and trigger, what they cost in context, and whether their docs match their code. Read the Skill Vision guide.
  • Research Hound is a scientific research workflow for Claude Code: five stages and 21 runnable scripts that verify citations, audit claims and lint a manuscript.
  • Agent Oracle is an AI agent orchestration design skill. It decides how many agents a task needs and what stops each loop before any work starts.

The guide to Research Hound and Agent Oracle covers both.

LLM Dogfight

LLM Dogfight is a terminal UFO dogfight between two local LLMs. Two small open models command rival saucer teams on CUDA, Apple MLX or the CPU, write down what went wrong after every loss, and can evolve their doctrine with a genetic algorithm. It is the most fun way I know to watch small models make decisions under pressure. See the LLM Dogfight guide.

mr-usage

mr-usage is a desktop API credit monitor. One window shows what you have left at 44 paid APIs, from Firecrawl and SerpApi to OpenRouter, ElevenLabs, RunPod and Ahrefs. Your keys stay in your operating system's credential store. See the mr-usage guide.

TigerOCR

TigerOCR turns scans, photos and PDFs into clean Markdown with tables, LaTeX and figure crops. Claude, Gemini CLI, Codex and Cursor use it as an MCP server, and anything else can call the REST API. The research guides on this site were written from TigerOCR's output. See the TigerOCR guide.

For embedded developers: GlassVGL and LightEDA

GlassVGL is a frosted-glass theme for LVGL 9 on ESP32, STM32 and RP2040 boards, with free MIT sample projects. LightEDA is a desktop IDE for ESP32, Arduino and FPGA boards: build, flash, debug, simulate VHDL and Verilog, and design LVGL screens in one app. Both pages list their plans and prices.

Which one to try first

If you write Claude skills, start with Skill Vision and point it at your best skill. If you have an ESP32 and a car, start with OpenCarMind. If you just want to see two small models argue in your terminal, install LLM Dogfight and type dogfight.