Agents fly the legs.
You land the change.
Atlas ADE sets up Claude Code and Codex in any repository, keeps a generated chart of your code, and drives every change through one short loop. It stops only when it needs you, with the next command already on screen.
$ ade card add "greet takes a name" \ --path src/greet.py --done "greet(\"Ada\") says hello"c1 next: ade loop c1$ ade loop c1lane advance lane ~/.ade-work/61eadb80/c1build advance built with claudereview advance light review approvedtest advance passed: tests/test_greet.pyland done landed c8b7669b5c$
__ _ __| | ___ / _` |/ _` |/ _ \ | (_| | (_| | __/ \__,_|\__,_|\___| a small agentic development environment What can I do? Start ade init start a project here, or install into this repo ade link make a plain `ade` work in new shells Work ade card add <title> a change to make, with its paths and done line ade loop <card> lane, build, review, test, land ade conduct loop every open card, waking a model only when needed ade land <card> merge a reviewed, tested lane ade push send trunk and records; fast-forward only Look ade cockpit every open card and its next command ade status the one line above ade watch <card> follow a lane live ade telemetry tokens, cost and week meters ade budget lines, hooks and refusals against their limits Decide ade decide rulings, and questions with a default and a deadline ade observe a dated observation, rechecked later Tools ade handover --role <r> an agent's starting context ade q store long output, then page or query it ade session run a verb detached, list or stop it ade wait wait for a condition, with a bound ade graph regenerate the code graph ade rehearse full loops on throwaway repos, spending nothing ade uninstall remove what init wrote, keep your records ade <verb> -h for one verb's options · ade -h for the full list
Scripts fly the route. A model is on call.
Routine legs are scripts and wake nobody. A model is woken to build, to review, and when a stop needs judgement, and you only when the call is yours. Any stop goes into a hold, and the card rejoins the course where it left it.
- 01 · pickThe next cardThe conductor takes the oldest open card, or you name one, and notes anything that moved under its paths since it was written.
- 02 · laneIts own worktreeEach card gets an isolated git worktree, so builders never collide on your checkout.
- 03 · buildEither engineClaude Code or Codex makes the smallest change that meets the card's done line.
- 04 · reviewA patch, not a listNone for prose, one light round for code. The reviewer hands back a fix, applied once.
- 05 · testWhat the change reachesTests are picked from the graph. An unknown path runs a small critical set, never everything.
- 06 · landOne verdictMerge, one outcome record, and any reds carried to a single residuals card.
- 07 · holdRecover and rejoinEvery stop is routed by one table: retry a blip, wait on a condition, wake a model to judge, or park it for you. Rerunning resumes where it stopped.
Think of a small aircraft.
Every part of Atlas ADE has one job, the way every instrument on a panel does. Nothing flies the aircraft except the loop.
Agents start with context, not a blank slate.
The first time an agent touches a file, it is told who calls it, which tests reach it, which decisions govern it and the helper it should reuse, straight from the graph. Each briefing is capped at 250 tokens; across ADE's own code they measure 34 to 144, about 100 typical. A role's whole handover runs about 200 tokens on a new project and 550 on ADE itself, against a 20,000 budget.
# ADE handover: builder Builds one card in its lane: the smallest change that meets the card's done line. ## Open cards - c1 greet takes a name [mp/greet.py, tests/test_greet.py] state: not started; next: `ade loop c1` - c2 log every greeting [mp/greet.py] done: greet writes one line per call state: not started; next: `ade loop c2` ## Graph (4 files; each edit is briefed from it) - mp/greet.py: the starter's one function (1 callers) ## Budgets core lines 14 / 20000 ok hooks 4 / 5 ok refusal codes 36 / 50 ok ~193 tokens (budget 20000) · empty sections and c1's done line trimmed
Small on purpose.
Budgets are the only standing gates, and every merge prints where the kit stands against them. Your telemetry stays on your machine.
PLACARDLIMIT
Core code~20,000 lines
Warm-up per role~20,000 tokens
Hooks≤ 5
Refusal codes, each with a cure≤ ~50
Whole test suite< 5 min
Get Atlas ADE
Start a new project on it in an empty folder, or install it into a repository you already have. One command either way, with a dry-run diff first and a clean uninstall.
$ ade init --dry-run $ ade init $ ade loop c1
- The
adeCLI for macOS, Windows and Linux - Claude Code and Codex from one roster
- Python and TypeScript starters, Orca or CLI-only cockpit
- The kit as it is when you buy it
- Everything in the kit
- Every kit release for 12 months
- One-off payment, no subscription
Bought the £14.99 kit? Add 12 months of updates for £12.99 from your account, whenever you like. VAT and sales tax are worked out at checkout.
Questions
Does anything leave my machine?
Not through ADE: it has no network code, and its usage records and telemetry stay in your project. The engines are another matter. Claude Code and Codex send your prompts and the code they read to Anthropic and OpenAI under your own account, and each has its own telemetry you can turn off. Read Claude Code's data usage and Codex's settings, or our summary.
Which engines does it drive?
Claude Code and Codex, from one roster. Agent definitions are generated for each from the same source, and ADE starts the real claude and codex binaries on your PATH.
Can I try it without spending usage?
Yes. Set ADE_PROVIDER=fake and every loop runs end to end with a fake engine. ade rehearse does one full loop per config option on throwaway repositories.
Do I need Orca?
No. Orca is an optional display that also carries alerts to your phone. The CLI-only mode has every instrument.
Will it change my global settings?
No. The committed settings hold permissions only. Machine-specific hooks go in a gitignored local file, and ade uninstall removes everything init generated.
What does it need?
Claude Code or Codex, signed in with your own account, and git on macOS or Linux. Each download carries its own Python (and git, on Windows), so nothing else is installed. The TypeScript starter also needs Node 22.6 or newer. No tmux and no background services: long runs detach with ade session on every platform. Install guide.