Your AI setup works. Your team's doesn't.
magster is the team harness for AI engineering
Two years on real client projects, opening up: 12 tools, 6 video courses, and 2 books.
AI adoption is a team problem.
Every developer on the team built an AI setup they were proud of. Then we put them on one project. The harnesses fought — one dev's rules rewrote another's code, agents with different memory made opposite calls in the same repo, and every pull request turned into a negotiation. Different teams need different rules, too: a backend team and a mobile team don't review code the same way. The harness has to live per team and per project. Nothing out there shipped that way. So we built it.
What magster is
magster is the system we ended up with. We built tools where the gaps were and wired them together with the tools that survived our testing. The result is one harness, shared by the team and configured per project. We change something in it almost every day. When a tool stops earning its place, it goes. We won't promise you'll get eleven times faster — but on our own projects the team got measurably faster, and the frustration went away. The tools stopped fighting each other. So did we.
What's in the box
Twelve tools — everything we use, the way we use it.
claudish
Run Claude Code with any model, from any provider, or several at once. One harness, no lock-in.
mnemex
claudememEight layers of agentic memory, shared by the whole team across git and your codebases. Your agents remember what the team already learned.
claudeup
The harness installer. One unified setup per team and per project: plugins, skills, aliases, profiles, tools. Everyone starts from the same page.
magus
A set of plugins that make Claude Code, Codex, and Antigravity collaborate on the same work instead of working alone.
models.madappgang.com
A live registry of models and providers: prices, capabilities, what's actually newest, and what we recommend per task. Your agents read it as ground truth.
madbench
Our benchmarking system, built on promptfoo. When models disagree, we measure instead of arguing.
mcp toolkit
A kit for building your own MCP tools when the one your team needs doesn't exist.
claudeup index
A curated index of popular plugins and systems. We run each one inside our real harness and keep only what works.
mastra
A toolset for prototyping custom agent systems, built on Mastra AI. Try the idea before you commit to building it properly.
meroku
Heroku-style deployment for your agentic infrastructure, on your own AWS. The simplest way we found to run agents in production.
passflow
usepassflow.comOpen-source authentication and identity for people and for agents. We couldn't find one that treated agents as first-class users, so we built it.
passgate
Shared model access for the whole team, without issuing an API key to every developer. Works like a gateway, with no central server to babysit.
Learn it the way we did
Six video courses ship with the toolset. They start at team 101 — the basics of working inside a shared harness — and go up to a team-lead and CTO track: how to roll the harness out and keep ten developers and their agents on the same page. One course we think nobody else teaches: connecting your business team to the same harness with Claude Cowork. Marketing, sales, operations — the people who never open a terminal get the same agentic infrastructure as your engineers.
And two books
The courses teach the tools. The two books are about the culture of agentic engineering — how a team thinks and works when agents do half the typing. Every week there's a model that changes everything. By Friday it changes back. The tools will be different next year. The culture is what you keep.
Why we're opening it
We've taken this as far as one company can. Inside MadAppGang the harness gets tested by our teams, on our kinds of projects. A community will break it in ways we can't predict — and every fix makes everyone's harness better, ours included. Join early and you get access to all the tools and the first lessons, and a real say in where this goes.