Overview

Published 27/10/2026

Atlas ADE is a small, portable agentic development environment. You install it once, run ade init in a project, and from then on Claude Code and Codex work inside that project with the same settings, the same context and the same short loop for every change.

You decide what gets built. ADE does the routine work around each change (a clean workspace, the build, a review, the right tests, the merge) and stops only when something needs your judgement, with the next command already on screen.

Think of a small aircraft#

• You are the pilot. You choose where to go (cards), answer the radio when the aircraft asks (a decision or a stop that needs judgement), and land.
• The loop is the autopilot. It flies the routine legs of each change (pick, lane, build, review, test, land) as scripts, and when something stops it, routes the stop the same way every time.
• A model is the first officer, on call. The conductor wakes one only when a stop needs judgement; the flight then carries on. You are asked only when the call is yours.
• Claude Code and Codex are the engines. Interchangeable, each bolted on through the same mount. Either can build or review.
• The graph is the chart. It is generated from your code, imports and records, never drawn by hand, so it cannot drift from the ground it describes.
• Records are the flight recorder. Cards, decisions and outcomes are written by ADE's own commands, append-only, and travel with the repository on their own git branch.
• The handover is the pre-flight briefing. Any agent, starting cold or picking up after its context was compacted, gets the same short briefing from one command.
• The cockpit is the CLI, or Orca if you like. Every instrument is readable from the terminal. Orca, an optional desktop and phone app, can show the same flight and carry alerts to your phone.
• Refusals are stall warnings. There are few of them, they are loud, and each one says how to recover.

What it does#

• Sets up a project in one command. Claude Code settings with conservative permissions, hooks, agent definitions for Claude Code and Codex from one roster, a status line. A dry run shows the diff first, running it twice changes nothing, and ade uninstall removes everything it generated.
• Keeps a board of cards. A card is a short piece of work: a title, the paths it may touch, and a line that says what finished looks like.
• Briefs agents as they work. The first time an agent edits a file, it is told who calls that file, which tests reach it, which decisions govern it and which helper it should reuse, in at most 250 tokens (about 100 is typical).
• Lands every change through one loop. Each card gets its own git worktree, is built by one engine, reviewed, tested against what the change actually reaches, merged and recorded.
• Keeps your usage visible. Tokens, cost and time per card and per engine, recorded on your machine.

What it is not#

• Not a hosted service. ADE runs on your machine and has no network code of its own.
• Not a new agent. It drives the Claude Code and Codex you already use, with your own accounts.
• Not a framework your code depends on. Your project's code never imports ADE, and uninstalling it leaves your code as it was.

Next: install it, then land your first change.