Arthur Loop, in plain words
Arthur Loop keeps AI coding agents working for hours without you babysitting them, and without them shipping anything you did not approve.
You pick a master agent (Claude Code or Codex today), point it at your projects, and the loop takes over the boring part: asking the advisor for a plan, queueing the work, handing packets to the executor, holding results at a review gate, and stopping to ask you when a real decision comes up. Everything lands in plain files you can read, so a crash never loses state and no answer is ever acted on before it is saved.
While it runs, a resource governor watches the subscription quota your agents burn and the memory and CPU your machine has left. When the quota touches your reserve line, the loop checkpoints instead of burning the last of it.
Start here
curl -fsSL https://raw.githubusercontent.com/myrrazor/arthur-loop/main/install.sh | sh
arthur init --demo # a seeded loop you can poke at
arthur web # the console, at http://127.0.0.1:7433
arthur status # the same state, in your terminal
The map
| Page | What it covers |
|---|---|
| Install | pipx, uv, curl, the Windows installer, updating, uninstalling |
| Your first loop | arthur init, flow presets, picking the master agent, the demo loop |
| How the loop works | Advisor, queue, executor, gates, human decisions, and the files under it all |
| The web console | The loop canvas, quota and machine readouts, editing prompts, what it deliberately cannot do |
| Agents & adapters | The master agent, the six detected agent CLIs, advisors, executors, trackers, quota sources |
| Prompts | The three preset tiers, per-project overrides, and what edge prompts do today |
| CLI & MCP | The --json contract, tick verdicts, exit codes, and the six MCP tools |
| FAQ & Windows | Data, privacy, platforms, and exactly how experimental Windows is |
Everything here describes v0.2.x. The deeper reference material ships in the repo under docs/, including the adapter contract and the workflow runbook.