Agents, projects, workflows, a living Vault and signed proof — on your machine, connected to the same account you use on the web and in the terminal.
-2dd4bf?style=flat-square&logo=github){kind=icon}
AetherCloud is where the work actually happens: teams of agents read your files, build multi-step results, and drop them back into your Vault — while you watch, or while you sleep. Every file operation and every model output is signed into a chain of custody you can read, search and export.
AetherCloud is in beta. It's used every day and updated constantly, so things will move under you sometimes. If something breaks, tell us — issues here, or security reports through SECURITY.md.
One account. One balance. One memory. Pick the surface that fits the moment.
- AetherCloud Desktop — the full app: agent teams, the Coder IDE, workflows, Vault, Voice, proof logs.
- Aether Online — the hub: projects, repositories, memory, connectors, collaborators, logs, Actions. Home of the Aether Project Runtime.
- Chat · Code · Design — the web. Research and hand off to agents, code in the browser with the same backend as the desktop, or work on a visual canvas.
- Aether Agent — the open-source coding agent for your terminal. Hosted models or your own local Ollama.
- Aether Browser — a real Chrome your agent drives, with a live view you can take over. Built into Aether Agent, runs on Aether Cloud.
Aether Online — the hub. Everything above talks to the same Aether API and writes to the same Vault and memory.
AetherCloud is a native Windows app. It is not a chat window with a file picker.
| 🤖 Agent teams | Build agents in the Agent Studio — name, role, model — group them into teams, deploy the team on a task. Teams use MCP servers, a skills library, and a sandboxed browser when a task needs the live web. Neo plans and fans out, Kronus audits and repairs, Aether-Vision chains image and video models from one prompt. |
| 🗂 Projects + Coder IDE | Open a project and work next to the agents on one surface: browse the tree, run builds, review diffs, let agents edit alongside you. Git, terminal, a media viewport, and Local Actions built in. |
| ⚙️ Workflows | Chain agents into a repeatable pipeline — draft → build → verify → file — then schedule it in plain English ("every weekday at 9am"). It runs unattended and lands signed output in your Vault. |
| 🗄 The Vault | The working memory your agents reason from: your files, your past work, your decisions, rendered as a searchable galaxy. It syncs to Aether Online, so the same context is in your browser and your terminal. Your files never touch a third-party cloud. |
| 🎙 Voice | Talk to the same agent you type to. Aether Voice. |
| 🧾 Proof logs | Read and export the signed record of everything that happened. Provenance. |
| 🧠 Memory | Let the system learn how you work. Agentic memory. |
Aether Agent is the open-source (Apache-2.0) coding agent that bridges into the same platform. It reads your repository, makes the change, runs the checks you name, and shows you the exit code. Hosted models through your Aether account, or local Ollama with no account.
Install → launch →
/help → /models. That's the whole first run.
npm install -g aether-agents@latest --ignore-scripts
aether auth login
aether- Proof, not claims.
--test-cmdties "done" to a real command and a real exit code. - The effort dial.
/effortfromLOWtoCODEPRO— the same dial the desktop uses. - Project Memory.
aether -mcommits a versioned memory snapshot of your repo, linked to the Git commit, and pushes it to your project on Aether Online. - Review → ship.
aether reviewpicks what goes in,aether shipopens the pull request — and never publishes without naming the action.
When an agent hits a login, a 2FA prompt, or a portal with no API, it gets a
real Chrome, driven over MCP, with a live view you can watch from any device
and take over the moment a human is needed. The agent pauses, you log in,
the agent resumes. It runs on Aether Cloud (or self-hosted), it's built into
Aether Agent's remote-session (/rc) flow, and it ships standalone as
agent-browser
(npm i aether-browser · pip install aether-browser).
Aether Online is the hub. It is also where the newest piece of the platform lives.
APR is the runtime under every Aether project. One project identity crosses every surface; the project carries its graph (persistent knowledge), its DAG (active execution state), and its proof.
you → Aether Online → Aether Project → Aether API → the right capability
- GitHub integration. Connect a repository and the project follows it — branches, pull requests, checks and deployments show up in the hub, and Actions results land back on GitHub as checks.
- MCP routing. Aether is an MCP server. Point Claude, GPT / Codex, Hermes or any MCP client at it and those agents can drive Aether models, project memory and media tools directly — other agents become connectors into your project, not separate silos.
- The Aether API. The single programmatic entry: the model fleet, the memory graph, project memory (
status / get / stage / commit / open), media generation and the asset library, development context and lanes, and bounded Supercluster runs — plan first, then dispatch with scoped, one-use authority. - Supercluster, behind the curtain. When an objective needs parallel work, Aether organizes a Supercluster of agents for you. You never configure workers or a DAG by hand.
Every run has an effort dial. It moves how many phases run, how many sub-agents fan out, how many repair passes happen, and the UVT ceiling for the run.
CodePro is different in kind, not just in size. It switches on two things:
- Unlimited context. The run gets a dedicated space — a large overpool with quantized recall and a context chain — so a long session never loses its own investigation past the model's window. That's Unlimited Context, Aether's open-source engine, doing the heavy lifting.
- The System-2 compiler. Instead of forty agents with forty private scratchpads, CodePro keeps one live, shared transcript — a compact intermediate representation every agent reads and writes. A 40-agent workflow is 40 agents, one compiled mind, orchestrated. The answer still comes back as plain prose.
CodePro on SWE-bench — same model, same tasks, official harness. Only the memory layer differs.
| Instances | Model alone | CodePro | Lift |
|---|---|---|---|
| 27 | 3 (11.1%) | 7 (25.9%) | 2.33× |
| 89 | 6 (6.7%) | 16 (18.0%) | 2.67× |
| 180 | 14 (7.7%) | 27 (15.0%) | 1.93× |
SWE-bench-lite, scored by the official Docker harness. Model fixed at gpt-4.1-mini — the engine's value is clearest on a base model that forgets. The defensible large-sample claim is ~2× more real bugs fixed, and CodePro led at every checkpoint.
One fleet across every surface. Models unlock by tier; what your account can
reach right now is whatever aether models prints while signed in, or the
picker in the app.
| Tier | Unlocks |
|---|---|
| Free | Claude Haiku 4.5 · DeepSeek V4 Flash · Gemma 4 31B · one image teaser · 1 active project |
| Solo | + Claude Sonnet 5 · GPT-5.4 mini · Neo orchestrator · every image model |
| Pro | + Claude Fable 5.1 · GPT-6 Astra · Gemini 3.8 Flash · Claude Opus 4.8 / Opus 5 · GPT-5.5 · GPT-5.6 Sol / Terra / Luna · DeepSeek V4 Pro · Kimi K2.6 / K3 · Qwen 3.8 Max · Grok 4.6 · Gemini 3.6 Flash · video generation · Kronus · Aether-Vision |
| Team | Everything in Pro, pooled across seats, 10 concurrent runs, 64k output, ~2× the video caps |
Image, video and orchestrator catalogue
Image: Nano Banana / Nano Banana 2 / Nano Banana Pro · FLUX.2 Klein / Pro / Flex / Max · Recraft V3 / V4 · Seedream 4.5 · Riverflow V2 Fast / Pro · GPT-5 Image / Image Mini · GPT Image 2
Video: Seedance 1.5 Pro / 2.0 / 2.0 Fast · Veo 3.1 / Fast / Lite · Kling 3.0 Standard / Pro / Video O1 · Sora 2 Pro · Wan 2.6 / 2.7 · Hailuo 2.3 · Grok Imagine Video · HappyHorse 1.0
Orchestrators: Neo plans, decomposes, fans out to workers and assembles the result. Kronus runs a deep audit chain — scan, cross-reference, find, fix. Aether-Vision pairs a reasoning brain with the vision fleet so one prompt plans, prompt-engineers and renders image → video in a single run.
UVT is Aether's universal compute credit — the billing system behind every product. One balance, shared across the desktop app, Aether Online, Chat, Code, Design and the terminal. You see the cost of a thing as it happens, not on an invoice later.
A subscription includes a monthly pool; top-ups go beyond it on demand and apply instantly everywhere. Media is priced per image or per second of video, scaled to the vendor's real cost. Per-model monthly caps sit on top of the plan pool so one expensive model can't drain a month in an afternoon. Full rate card: PRICING.md.
Actions is how work gets verified and shipped. Connect GitHub in Aether Online, pick a repository, and every push or pull request can trigger a run. The result shows up in Recent runs with the UVT it captured and released — and on the pull request as a check.
There are two products behind that one decision:
Build & Test runs your project's configured checks — tests, lint, typecheck, build. Run it on your device through Local Actions in the desktop app for 0 UVT, or on Aether Cloud with an exact UVT quote and your explicit authorization before anything executes. Either way, the coding agent can drive the loop itself: run → read failures → fix → re-run → green.
Predator Security is a separate security CI, hosted-only. It verifies the exact commit against five locked assurance routes — Access Control, File Boundaries, Command Safety, Network Request Safety, Protected AI & Billing Authorization — and a clean run issues a Predator Security Certificate for that commit. The operation, checks and price are locked and can't be edited, so a certificate means the same thing for every repo.
Actions in Aether Online — one click for Build & Test, one for Predator Security.
Local Actions — the .aether-ci.yml
Drop this in the repo root. Checks run in the app sandbox on your machine — no
cloud compute, no metering. Three gates, any subset: block git commit, block
git push, or expose a read-only ci_run tool to the coding agent. You approve
the exact commands once; the agent can never self-approve.
version: 1
gates:
commit: false
push: true
agent: true
checks:
- name: tests
type: test
run: pytest -q
- name: lint
type: lint
run: ruff check .Every approval and every run is signed into the Protocol C chain. A signer
outage logs the run as UNSIGNED; it never blocks your local checks.
Most AI tools start from zero every session. AetherCloud doesn't. Memory lives in the cloud, keyed to you, and every surface reads the same store — teach it once anywhere, it knows everywhere.
- Facts are extracted from your sessions, scored for importance and confidence, and re-scored over time — episodic, semantic and behavioral, global or per-project.
- Skills (QOPC) — the Quantum Optimized Prompt Circuit watches what you accept, revise, publish or discard and tunes its own prompt weights. No config, no fine-tuning; it gets better the more you use it.
- Dream Logs cluster what was learned between sessions and write a short synthesis, so the system arrives already up to speed.
- Project Memory is a versioned snapshot of a repository, tied to an exact Git commit — commit, push, pull, diff, reconcile. Git, for memory.
The Memory Graph in Aether Online shows how your code, memory and Vault connect. Memory chips appear inline in chat so you can see what the agent is drawing on. Retrieved memory is evidence for the agent, never instructions.
Every file operation, agent run, model output, auth event and chat turn is hashed (SHA-256), signed with a one-shot key that's destroyed after use, timestamped (RFC 3161), and appended to an immutable log that chains to the entry before it. A tampered response fails its signature and the system refuses to proceed.
The log is viewable in the Logs tab and the Aether Online hub, searchable by event type, and exportable as an evidence-grade proof package you can hand to an auditor, a client or a court. The chain itself is open source — AetherCloud is built on it. Patent pending.
Two more things sit on top of it: ATLAS, a ground-truth engine between your models that keeps outputs anchored to verified reality, and Predator, whose security certificates are signed into the same chain.
Voice isn't a separate app. The mic in the composer opens a session that speaks into the same chat thread, runs the same agent — same tools, permissions, memory and billing — and renders into the same transcript. Typing and talking are two inputs to one conversation.
- Live conversational awareness. The surface you're on tells Voice what's on screen — the file you have open, the project you're in, the terms in front of you — so it stays in context instead of guessing.
- Barge-in. Start talking while it's speaking and it stops, drops the synthesis, cancels the stream and listens.
- Streaming speech. Replies are chunked and voiced as they arrive, in order, with markdown turned into something that sounds right out loud.
- Provider failover. If one voice provider stumbles, the next picks up mid-turn.
- Download — the installer provisions its own runtime (Node, Git, ripgrep). About a minute.
- Sign in with your Aether account — the same one for app.aethersystems.net and
aether auth login. - Open a project, deploy a team. Or just ask it something.
Want the terminal too? npm install -g aether-agents@latest --ignore-scripts
(or pipx install aether-agent), then aether auth login. Upgrade or top up
anytime at aethersystems.net/pricing.
This repository holds the AetherCloud product docs, release notes and media. It does not contain the application source. Security reports go to security@aethersystems.net — see SECURITY.md.
AetherCloud is proprietary. Aether Agent, Aether Browser, Unlimited Context and Protocol C are open source under Apache-2.0. Terms.
AetherCloud — Aether AI LLC · Proprietary · Patent Pending #64/010,131 · All Rights Reserved
Created by Brandon Barrante · founder, Aether AI
Download · Product page · Aether Online · Pricing · Release notes · Terms · Security