v0.2.0 · prototype · open source · Windows
Your own AI team.
Chat tools answer one question at a time and forget. Aether runs a team of AI agents that plan, build, review and remember, under your rules, on the AI model you choose.
AI in three minutes
From chat to a team you can trust: four steps.
Most people meet AI as a chat box. That is step one of four. Each step adds something a business needs.
Chat
You ask, the AI answers. When the conversation ends, it forgets. It can't open your files or take an action.
Like a brilliant consultant on the phone, with no access to your office.
Agent
An AI given a goal, tools (files, search, code) and permission to take several steps on its own, checking its results as it goes.
The same consultant, now given a desk, a laptop and a task list.
Orchestration
Several agents with different roles (a lead, specialists, a reviewer) working on one piece of work, with someone coordinating who does what.
A project team with a project manager.
Harness
Everything around the AI that makes it safe and useful at work: memory, rules, approvals, budgets, records and the keys to outside services.
The AI model is the engine. The harness is the car, the dashboard and the rules of the road.
Most organizations stop at step 1. The value is in steps 2 to 4, and step 4 is where control lives.
Why own your harness
Rent the engine if you like. Own the car.
AI models are improving and changing price every few months. The harness is what stays: your rules, your records, and what your AI team has learned.
Your data stays yours
It runs on your machines, with your keys, and can use AI models that run entirely on your own hardware for sensitive work.
No lock-in
Swap or mix AI providers as prices and capabilities change. The work and the memory stay.
Knowledge that stays
What your AI team learns about your work stays with you, not with a vendor.
Accountability
A record of who decided what, which agent did what, and with which files: the basis for audit and compliance.
Cost control
A budget for each piece of work, and a stop when it is reached.
Your process, not a vendor's
Your roles, your review steps, your approval points.
What Aether adds
A harness with a working team already inside.
A team that remembers
Each agent keeps its own memory and decides what is worth keeping. Next time, it brings the lesson with it.
Proof · A lead applied a lesson from its previous project without being told, and said it came from its memory.
A built-in second opinion
An independent reviewer checks the work before you see it.
Proof · Two projects in a row were reviewed, then graded 27 of 27 and 29 of 29 by checklists the team never saw.
You stay in charge
You approve the plan once and accept the result. Spending and keys stay under your authority.
Proof · After a live run, the AI key appeared in none of the 1,252 files checked.
Every step on record
Every message and file is kept and can be inspected afterwards.
Proof · Delivered files are kept with a fingerprint, so any later change shows.
Built to keep working
It picks up after a crash, and backup and restore keep everything.
Proof · Backup, restore and reinstall kept all 191 data files byte-identical.
Your choice of AI
Claude, OpenAI, DeepSeek, or models on your own hardware (LM Studio, llama.cpp), chosen per agent.
Proof · Live runs on DeepSeek and LM Studio. Claude check scheduled.
Open source (MIT). Anyone can inspect it, extend it and run it. See the evidence →
See it
Nothing hidden: the memory, the team and what they've learned are all on screen.
Memory you can see
Each lead agent has its own Memory Fabric: lessons, warnings and notes it chose to keep, linked to each other. The viewer shows what it remembers, how items connect, and when each was recorded. Memories are never deleted; the ones an agent stops using sink out of view. A seal shows the memory hasn't been tampered with. The view is read-only: the agent manages its own memory.
A team you can staff
Leads plan the work and hire staff; staff do the tasks. Each has a role, a specialty and a home of its own. Aether ships with a starter team of specialists, and you can create your own.
Skills you can teach
A Teach Frame captures an approach: when it fits, how to do it, and how to check it. Any agent you recommend it for can use it, and Aether ships a School of frames for its staff. When two specialists who had drawn a poor cat were given two construction frames, the next two animated drawings were each passable on the first attempt.
How a piece of work runs
Five steps, and you decide at two of them.
Brief
Say what you want delivered, why, and what you'll spend.
Team
The lead proposes a plan and the specialists it needs.
Approve you
You approve the plan, the team and the budget, once.
Work
Agents deliver, review each other's work and revise.
Accept you
You accept the result. Each agent files what it learned, for next time.
What you could use it for
Work that benefits from a second pair of eyes and a memory.
A report from source documents
One agent drafts, another checks it against the sources, and you get both the report and the review.
A code change
A developer agent makes it, a reviewer checks it, and the record shows what changed and why.
Research that builds on itself
Each brief starts from what the team learned in the last one.
These are the kinds of work Aether is built for. The demonstrations so far are listed in the evidence below.
Where we are
An early release, tested honestly.
Works today
A Windows app with Python included. Install, sign in, connect an AI provider, chat with an agent that has memory, and run planned work with review and acceptance. Installed and used end to end on a second computer.
Next
More AI providers checked live, more operating systems, and broader independent review.
Limits
This is not a security sandbox: code an agent runs, runs with your user rights, so give agents work you would run yourself. Results so far come from small numbers of recorded runs, not a large-scale study.
For tech leads
Under the hood.
- Mission Control. A web app on your own machine, and the same actions through an HTTP API. People can use it, and so can their AI assistants.
- The work loop. A lead plans a chain of tasks, and specialists close or block their own. Code enforces the hard rules: declared output files are checked before dependent work starts, written answers need a review tied to the exact bytes, and a budget stop is a stop.
- Memory. Each lead has a Memory Fabric: notes, links and lessons it writes, recalls and maintains itself, during the work and at its close. The viewer shows what it holds and when each item was recorded.
- Models. One OpenAI-compatible route for every provider, with a model chosen per agent. Model settings stay separate from the agents' identities and memory.
- Keys. Stored encrypted, granted per account, and never handed to code an agent runs.
- Recovery. Work resumes from durable records without repeating an action already recorded.
Frameworks give you parts. Aether is the assembled thing.
Agent frameworks are good tools for developers who build agent systems, and much of what Aether does can be built with them. Aether is the assembled application, with these working together by default.
Memory
With a frameworkMemory stores you wire in and scope yourself.
In AetherEach lead owns its memory, decides what to keep, and says when a decision came from it.
Review
With a frameworkA reviewer step you add and define.
In AetherAn independent reviewer in every project, before acceptance.
Human control
With a frameworkInterrupts you place in your workflow.
In AetherOne approval per plan, acceptance of the result, budget stops.
Interruption
With a frameworkCheckpoints; you make each step safe to repeat.
In AetherResumes from durable records without repeating a recorded action.
Evidence
With a frameworkTracing you add.
In AetherEvery message and delivered file recorded, with fingerprints.
Packaging
With a frameworkA library you run as your own service.
In AetherA Windows app with installer, accounts, backup, restore and uninstall.
A fair comparison: the framework column describes what the frameworks' own documentation offers, not what can't be built with them. A capable team could build Aether on one. The difference is the time it takes, and defaults that have each been shown in a recorded run.
The evidence
Every claim on this page, with its date and its limits.
| Claim | What was run | Date | AI provider | Limits |
|---|---|---|---|---|
| Two projects planned, built, reviewed and accepted | The full app, with the operator's words fixed in advance; 73 model calls | Oct 28 | DeepSeek | Two projects; earlier version of the app |
| Graded 27/27 and 29/29 | Hidden checklists the team never saw | Oct 28 | DeepSeek | Two projects; earlier version |
| A lesson applied from memory, unprompted | The second project's lead loaded and used a lesson it filed in the first | Oct 28 | DeepSeek | One observed case; earlier version |
| Picks up after a crash | Server stopped hard between agent turns; signing in resumed the work, and nothing was done twice | late Oct | DeepSeek | Between turns; earlier version |
| Backup and restore keep everything | Backup, verify, restore, uninstall and reinstall: 191 data files byte-identical | Oct 28 | n/a | One data set; earlier version |
| Keys kept out | Key scan of 1,252 files after the run | Oct 28 | DeepSeek | That run's files; earlier version |
| Planned work with checks before acceptance | Two requests through the work loop, about $0.35 in total | Nov 1 | DeepSeek | Two requests |
| Taught specialists, first attempt | Two animated drawings, each passable on its first submission after construction lessons | Nov 5 | DeepSeek | Two trials |
| Installed and used end to end | Install, sign-in, chat, a mission and two pieces of work, download, on a second computer | Nov 7 | DeepSeek | One computer, one operator |
| Chat with memory across conversations | Three answers and two closings in the agent's own words; a fresh conversation recalled the facts and named the previous session; 21 requests | Nov 8 | DeepSeek | One run |
| The same check on Claude | Scheduled | — | Claude | Not yet run |
Get Aether
Free, open source, and yours to run.
- Download the Windows package from GitHub.
- Unzip it, double-click
Install.cmd, thenAether.cmd. - Sign in with
admin/aether-admin. The welcome asks you to set your own password, choose an AI provider and paste its key.
Or let your AI assistant install it. Open the unzipped folder in your AI assistant and ask it to follow INSTALL_WITH_AI.md. It does every step except your password and your key.
Needs: Windows 10 or 11, and an account with an AI provider, or a model running on your own machine (LM Studio or llama.cpp).
License: MIT · Report an issue on GitHub · aethercorecontact@gmail.com