AGMX

Hand work to a coding agent the way you hand it to a person: the agent runs on your own machine, under your own CLI subscription, a different agent reviews what it did, and the system merges and pushes. Your code never reaches the server.

If you are the one running the team

There is a screen for you at https://dev-test.agmxcode.com/dashboard. You do not need any of the wiring below to open it, and no CLI: sign in on this server and it shows your team's agent roster and what each agent is doing right now, the queue and — in plain words, not just a number — why anything in it is sitting still, the machines that take work and when each last checked in, the gates waiting on a person to decide, and the cost brakes with how close each one is to stopping work.

The page itself opens for anyone; nothing real appears on it until you sign in, and where your team signs in through its own identity provider the dashboard sends you there to do it. One panel animates the flow of work through the system rather than reporting it — those figures are an illustration of the shape, never a measurement of your team.

The fastest way: let your own agent do it

Paste this line into the Claude Code, agy, codex, qwen, kimi or opencode session you already have open:

Read https://dev-test.agmxcode.com/install.md and connect this machine to AGMX following that guide.

The agent installs the package, walks you through the sign-in, wires MCP for whichever CLI it is running in, and then calls a tool to prove it worked. The only thing you do is confirm the sign-in in your browser.

Or do it yourself, two commands

pipx install agmx
agmx login --server https://dev-test.agmxcode.com

--server is not optional: agmx has no default server, because a built-in address is always some other team's install. There is no password prompt either — this server signs you in in your browser, or prints a device code on a machine that has no browser.

Signing in is the whole wiring step. agmx mints a key that lives 30 days and connects every coding-agent CLI it finds by itself — claude, agy, codex and opencode each through their own command, and ~/.agmx/mcp.json for qwen and ~/.kimi-code/mcp.json for kimi. There is nothing to copy and nothing to export. When the key expires, run the same command again: it reuses the session already stored, without a browser.

Then ask your agent: “call agmx's get_status”. A list of your team's work means you are connected.

No account yet?

AGMX is sold one install per team: an admin invites you by email and you sign in on this very server. There is no self-serve signup.

To let this machine take work

agmx runner add <project_id>=<path to that repo on this machine>
agmx runner install

Signing in only lets you give work; this is what makes it run here — in the background, coming back after a reboot. Get project_id from your team's project list and copy it exactly: an id that does not exist is accepted without complaint and then never receives a single run. agmx runner list shows what this machine is registered for.

Runs execute as you, with your privileges, under your own CLI subscriptions, in a throwaway git worktree of your own checkout. Your teammates' tasks are not accepted here unless you ask for --serve-team.