mega-edo/mega-tron

MEGA Tron is a local layer that sits above Codex, Claude Code, and Gemini CLI to manage the custom skills (also called slash commands or agents) you load into those tools. It addresses three problems that get worse as you accumulate more skills. The first problem is token waste. When you type even a simple message, the host CLI injects a large block of skill metadata into the request, regardless of whether those skills are relevant. MEGA Tron replaces that fixed injection with a per-turn semantic router: your prompt is embedded and ranked against every skill in your pool, and only the top-K relevant ones are sent along. According to the project's benchmarks, this reduces catalog token cost from roughly 1,200 to 3,500 tokens per turn (native hosts) down to about 100 to 600 tokens, while maintaining 0.96 coverage of the skills that were actually needed. The second problem is host isolation. If you maintain skills in Codex, you have to separately maintain them in Claude Code and Gemini CLI. MEGA Tron stores all skills in a single master pool directory and creates symlinks from each host's skill folder to that pool. Editing a skill in one place updates it everywhere. The third problem is that none of the three hosts records whether a skill actually helped. MEGA Tron captures a verdict for every skill use, labeling it HELPFUL, HARMFUL, or NEUTRAL, along with the prompt context and which host ran it. These verdicts feed back into ranking: a skill that fails repeatedly across sessions gets auto-archived, a skill that performs well in Claude lifts its rank in Codex when a similar prompt arrives. A local dashboard web interface shows which skills are performing, which have broken silently, and how performance changes over time. The easiest installation path is to tell your AI coding assistant to follow the agent installation guide linked in the repository. The tool requires Python 3.11 or later and is licensed under Apache 2.0.