Start with a plan that can be checked.
Your design becomes a compiled, hash-bound plan. Milestones say what to build, what evidence to keep, and when to ask you. You approve the exact plan before implementation starts.
A local-first OS for your coding agents
Build hard things.
Build them well.
Give your coding agents a design doc. aidashos keeps their plans, attempts and reviews in a durable ledger on your computer.
AGPL-3.0-or-later · Apple Silicon macOS · Developer preview
YOUR DESIGN DOCUMENTBuild something
worth finishing.
worth finishing.
COMPILED PLANHash-bound
One local, durable ledgerPlan Attempts Artifacts Decisions
01Compile / A compiled plan
Free developer previewBring your subscriptionsKeep a local record
01 / The plan is the center
A little structure.
A lot less guesswork.
For a small task, your agent can work in one pass.
For work with dependencies, reviews, or recovery, give it a plan.
Give each attempt a bounded job.
The senior agent works in an isolated Git worktree. Its patch and test evidence become durable artifacts, ready for the next stage.
A separate session checks the work.
The staff agent reviews the actual change against the plan. Review can send work back for a revision before you consider integration.
Keep the consequential decision yours.
You inspect the reviewed result and approve integration. The local ledger retains the plan, attempts, artifacts, and decisions.
THE ONLY NEW TERM
A CGD is a design document with the sections aidashos needs to compile it.
Use the GAWD template, or write your own document that meets the same contract.
02 / Three roles, one shared record
Build. Review.
Keep an eye on it.
Two frontier agents and a local junior work from the same plan.
Their roles are different. Their evidence stays together.
01Frontier model
Senior
Builds the change
Implements a bounded task and records the result for review.
02Frontier model
Staff
Reviews independently
Checks the actual work in a separate session against the plan.
03Local model
Junior
Watches the process
Assesses stalls and checkpoints so a stuck attempt can be surfaced.
Use the subscriptions you already have. Configure separate providers when both are available, and keep review in a separate session. Provider quotas and account billing settings still apply.
03 / A durable place to leave off
The session ends.
The record stays.
Plans, attempts, artifacts, reviews, and approvals live in your local ledger.
When a process stops, that evidence remains.
An interrupted agent may need a new attempt; your laptop still needs to be awake to do the work.
04 / Start with the AI you already use
Try the
developer preview.
The installer and guide are on GitHub and currently require access.
FOR USERS WITH GITHUB ACCESS
Read the GitHub install guide and help me install aidashos for this project.
Show me the setup plan before making changes.
You will need an Apple-Silicon Mac, Homebrew, Command Line Tools,
authenticated frontier CLIs, and a local model for the junior tier.
Provider sign-in and native verifier qualification need your participation.
Before you start
A few useful
details.
Do I need another chat app?
No.
Use your existing coding agent through MCP.
To inspect plans, evidence, or decisions visually, ask your agent to open the aidashos Cockpit.
Why three models?
Implementation, review, and process supervision need different jobs and contexts.
A separately trained reviewer can make different mistakes, but diversity does not guarantee correctness.
The local junior assesses progress; it is not a general hallucination detector.
Does everything stay on my machine?
The control plane and ledger run locally.
Your configured frontier CLIs still connect to their providers under your accounts.
There is no aidashos hosted backend.
Can I just let it run unattended?
Only within the plan's approved scope.
Some failures can be retried; uncertain effects and consequential decisions can require you.
Integration and deployment keep their operator gates.
Is this ready for everyone?
This is a developer preview for Apple-Silicon macOS.
Fresh-machine installation and whole-stack memory qualification are still unverified.
Start with a small project and inspect the result.
Read the open issues.
A better place for your next hard task