Introducing · MEGA Tron
The skill layer every coding agent deserves.
Open source. Unifies the skills your coding agents already use, hands them only the ones that match the task, and grades every call so the next one gets sharper.
Works natively with
One layer
Every coding agent.
One unified skill catalog.
Claude, Codex, and Gemini refract through MEGA Tron today, with Hermes joining next. What comes out the other side is a single catalog every agent can draw from.
→ One catalog, served to every agent
The problem
Skills got added.
The plumbing didn’t.
Every coding agent built its own skill system in isolation. The result is three problems that compound the more skills you accumulate.
Fragmentation
Every agent keeps skills in its own silo.
Skills authored for Claude don't show up in Codex. Same skill, three formats, three loaders.
Context bloat
More skills, worse skill use.
Catalogs get dumped wholesale into the prompt. The right skill is in there — the model just can't see it.
No feedback loop
Nobody measures what helped.
Skills run and disappear. No record of why, no signal on whether it worked, no way to get sharper.
How MEGA Tron works
Three moves. One layer.
MEGA Tron sits between your agents and their skills — unifying, filtering, and grading. Every call gets sharper than the last.
Unify the catalog
Author a skill once. Claude, Codex, and Gemini all see it on the very next turn. One source of truth, no per-agent forks.
Optimize per turn
MEGA Tron reads your prompt first, then ships only the skills that match. ~600 tokens per turn for skill context, no matter how the pool grows.
Evolve from every call
Each session grades itself at the end. Outcomes feed the next ranking; skills that consistently fail get retired automatically.
The dashboard
See what your skills are actually doing.
mega-tron dashboard opens a local web UI that surfaces every HELPFUL, HARMFUL, and NEUTRAL verdict your three CLIs recorded automatically, plus the ones you add by hand. Three tabs: context savings, skills overview, and a filterable human-in-the-loop verdict log.
mega-tron · skill observability
Tokens shipped per turn
Hover a row to focus
benchmarked at pool = 59 skills · README §📊
Each turn, every host injects its full skill metadata blob before reading your prompt. mega-tron rebuilds the catalog semantic top-k per turn, so the cost stays flat at 106 regardless of pool size.
Works with the agents you already use
Open source · Apache 2.0
Bring every skill into one router.
MEGA Tron is on GitHub. Wire it in once and Claude, Codex, and Gemini all draw from the same catalog, with grading built in.
View on GitHub