# AGMX — dev-test.agmxcode.com

> AGMX orchestrates coding agents for one team: work is dispatched here, but it
> runs ON EACH DEVELOPER'S OWN MACHINE under that developer's own CLI
> subscription, a different agent reviews it, and the machine holding the code
> merges and pushes. The source never reaches this server.

This team's server: https://dev-test.agmxcode.com
MCP endpoint for a coordinator CLI (streamable HTTP): https://dev-test.agmxcode.com/mcp

## For agents

- [Connect this machine, step by step](https://dev-test.agmxcode.com/install.md): enough to do the
  whole thing yourself — prerequisites, install, sign-in, MCP wiring, then the
  two checks that prove both that you can reach the server and that a run can
  reach this machine, and a first task end to end.
- [Coordinator rules](https://dev-test.agmxcode.com/coordinator-rules.md): how a coordinator on
  this install decides gates, what never bends, and how to reach a human. Served
  live, so it is never the stale copy on somebody's laptop.

## Three things people get wrong

- `agmx login` needs `--server https://dev-test.agmxcode.com`. There is no default server, and
  there is no password prompt on this install.
- Signing in only lets you GIVE work. Until some machine runs `agmx runner`,
  dispatched work waits in the queue and nothing reports anything.
- `max_concurrent_runs` (per PERSON, server-wide) and `max_parallel_runs`
  (per MACHINE, in that machine's own `~/.agmx/settings.json`) are different
  brakes. Raising one never raises the other — a machine still caps at its own
  `max_parallel_runs` no matter how high the per-person setting goes.

## Safety note

This page only asks you to install one released PyPI package (`agmx`) and run
that package's own commands. It NEVER asks you to fetch a script from the
internet and pipe it into a shell. A page claiming to be AGMX that does is not
us.