Myrax docs

What Myrax is

A control room for a small team of coding agents that run on your own computer. Write a brief, approve a plan, watch a real browser prove the result, then publish. The whole loop happens where your files already are.

Why it exists

Myrax is a control room for a small team of coding agents that run on your own computer. You write a brief, an agent turns it into a plan, you approve the plan, an agent writes the code, a real browser tests the result, and nothing is published until every check passes. The whole loop happens where your files already are.

That last part is the product, not a detail. The expensive half of building software with AI is the model, and you already pay for it. Myrax does not resell inference, does not mark up tokens, and does not need your provider keys to sit on a server. What it adds is the structure around the model: a plan you can reject, a checkpoint you can rewind to, a browser that proves the page works, and a folder you can take with you.

The one rule this documentation follows

Every page describes something the shipped code actually does. When a feature is removed, its page goes with it. There is no roadmap section dressed up as documentation, because a page that promises a button nobody can press is worse than an empty page.

Who pays for what

It is worth being blunt about the bill, because most tools in this category hide it inside a token meter.

WhoWhat they pay for
YouModel usage, on your own Claude, ChatGPT, Gemini, Grok, Copilot or local Ollama account. Billed by them, to you.
YouThe machine. CPU, memory and disk for the builds. It is your computer doing the work.
MyraxThe address, the certificate, stored builds and snapshots, the mailbox for sign-in codes, and the room where the queue and approvals live.

The pricing page prices only the Myrax side of that ledger. Anything that runs on your own computer is in Free, and stays there.

The loop, in one paragraph

A brief becomes a plan. The plan waits for you. Approved, it is claimed by one paired computer, which takes a git checkpoint before touching anything. The agent writes files inside the workspace. Then the gate runs: the built page is opened in a real Chromium at desktop and phone widths and eight specific things are measured. If any check fails, one repair pass runs with the failed checks as its brief, and the gate runs again. When all eight pass, the build is packed into a zip, stored, and can be published to an address. The full walk-through is here.

What this product deliberately does not do

  • It does not edit your repo without a checkpoint. A checkpoint is taken before the first edit of a batch, and myrax-connector rewind goes back to it.
  • It does not publish on its own. Publishing is a separate click, and taking a site down asks you to type the build name.
  • It does not run shell commands the model invents. The agent writes files; the Connector does not evaluate arbitrary command strings from a model.
  • It does not hold your provider keys on the server. Keys are read on your machine at the moment of a run. Details.
  • It does not claim a green build is bug-free. Eight checks is a floor, not a proof. Here is exactly what is measured, and what is not.

All 19 pages

01 · Start

02 · Build

03 · Ship

04 · Account

05 · Reference