A local-first workspace for coding agents, with you in control.
Plan the change. Coordinate the team. Inspect the evidence. Approve what ships.
Website · Downloads · Get started · Explore the workspace · Architecture · Contribute
Aether turns an objective into a reviewable implementation plan, then runs approved tasks in isolated Git worktrees. Coding leads can work in parallel, ask same-model helpers to review their changes, and integrate the returned findings. Plans, outputs, tests, diffs, prompts, skills, and project memory stay in a local SQLite store.
You approve the plan before execution and the final patch before it reaches your project. Local-first describes storage and orchestration; selected cloud providers still receive the task context sent to them.
The real desktop interface, captured with a disposable Orbit demo plan. No personal sessions are shown.
| Capability | What it does |
|---|---|
| Coding teams | A lead implements, 1–3 same-model read-only helpers review, and the lead integrates their findings. Bounded concurrency, cancellation, and visible handoffs. |
| Scoped execution | Task dependencies, file ownership, contention locks, isolated worktrees, retries, and serialized integration. |
| Live Graphify | Inspect actual source symbols and relationships with local AST extraction, search, and task-scope highlighting. |
| Code-OSS IDE | A separate VS Code source-based workspace with Explorer, Git, terminal, Open VSX/VSIX extensions, right-side Ava chat, and Back to Aether. |
| Searchable models | Search runtime catalogs and published aliases across Codex, Claude, Antigravity, and compatible APIs; availability remains explicit. |
| Context efficiency | Economy/Balanced/Thorough budgets, bounded conversation windows and compact review logs while retaining full local evidence. |
| Nexload prompt core | Pinned, compact reasoning profiles with inspectable source provenance, prompt snapshots, and bounded context. |
| Project skills | Discover skills from OpenAI, Anthropic, and Vercel; inspect a pinned package before installing it into your project. |
| Ava & desktop activity | A reactive companion and floating project widget show work, questions, approvals, and completed tasks. |
| Reviewable evidence | Actual command output, exit codes, diffs, ordered events, and separate plan/apply approvals. |
| Project continuity | Local history and memory, analysis-to-implementation follow-ups, automatic initial Git setup, and English/Persian/Arabic text support. |
Follow the handoff.Inspect helper findings and the evidence returned to the coding lead. |
Explore the code.Navigate source relationships without a hosted semantic indexing step. |
A small presence on your desktop
Filter by project, browse completed tasks, pin or collapse the widget, and open the relevant session. Widget actions navigate; they never approve or apply changes.
Banner: AI-assisted artwork using an actual app screenshot. Orbit views use demo data; the coding-team view comes from a real isolated Codex test. Image provenance and regeneration.
Open Settings for the connection controls. GitHub uses your existing official GitHub CLI sign-in (or opens its browser login), verifies the account and lists up to 50 repositories with a searchable clone action. Git and GitHub CLI must be installed; credentials stay with the CLI.
For Telegram, create a dedicated bot with @BotFather, enter its token in Settings, consent to activity sharing and send the generated /start command to the bot in your private DM. Confirm the displayed user in Aether. Notifications and /status / /pending work while the app is running and online. Optionally enable remote decisions: review the complete plan or patch first, then use its buttons to approve, apply, reject or stop. No commit or push is triggered by those buttons. See setup, safeguards and limits.
Downloaded desktop builds include native Codex and its execution helpers; use Settings to sign in on each device. Claude and Antigravity offer Install & sign in through official installers.
From source, tested on macOS arm64: you need Node.js 24+, npm and Git. Prepare the pinned Codex runtime below, or use a compatible official CLI installation. Rust is not required.
gh repo clone gecut/aether-agent
cd aether-agent
npm ci
npm run prepare:providers
npm run devOpen Settings, run provider diagnostics, and choose your planner and coding model. Leave a model identifier empty to use the runtime default, or enter an account-supported identifier. Use the provider's supported sign-in flow if needed.
Use Import project to clone a GitHub repository (owner/repository) or another HTTPS Git URL, or open an existing local project created with any tool. Choose a clone location, then optionally review and install Aether’s project workflow and feature verification skills before continuing to the workspace or IDE. Project actions → Project setup & skills reopens this step later. Private clones use an existing Git credential helper or signed-in GitHub CLI; Git must be installed. Import does not run dependencies, submodules or project scripts.
The home screen lists your recent projects. Use Continue in IDE to reopen the latest one, or choose another project from the sidebar picker. Aether remembers the last project opened across restarts; folder paths distinguish projects with the same name.
Open IDE → Set up full IDE to download the checksum-verified Code-OSS runtime. Install compatible extensions from Open VSX or VSIX in its Extensions view; Microsoft-only Marketplace extensions are not guaranteed. The lightweight Quick editor remains available without that download. See IDE details and measured context savings.
Every successful push publishes Windows x64, macOS Apple Silicon/Intel, and Linux x64 development builds to GitHub Releases, with SHA256 checksums. See automatic release details.
For a production build or a local macOS package:
npm run build
npm start
# Create an ad-hoc signed .app and ZIP in release/
npm run packageThe local app is release/mac-arm64/Aether Code.app. CI also packages Windows and Linux downloads. macOS previews are ad-hoc signed but not Apple-notarized; Windows builds are unsigned. See Mac download and approval notes. Automatic app installation is not implemented.
- Open a local project with the project switcher or ⌘ O. Aether initializes Git and creates a local initial baseline for eligible new folders; it preserves existing history and staged work.
- Describe the outcome. Inspect the generated tasks, dependencies, file scopes, model choices, and verification commands.
- Select Approve & run. Workers execute in isolated worktrees; overlapping scopes run sequentially. New plans enable one coding review helper by default; adjust the limit in Settings.
- Inspect Changes, tests, helper findings, and Trace. Completed analysis sessions can continue into a separate implementation plan.
- Select Apply changes after review. Aether checks your original repository still matches its clean baseline and applies the patch without publishing or committing it.
Execution needs a clean original working tree. Aether never stashes or resets your changes. Dependency installation is not automatic; see development and troubleshooting.
| Provider | Current capability |
|---|---|
| OpenAI Codex | Verified planning, coding teams, isolated implementation, streaming, cancellation, and read-only final review through the official CLI. Uses an existing supported CLI login. |
| Claude Code | Read-only planning/review adapter. Actual integration remains unverified; coding is disabled pending a verified sandbox. |
| Google Antigravity | Text-only chat/planning/review with a tool-free temporary profile, official OAuth setup and model discovery. Coding remains disabled; the tested account/location was blocked by provider eligibility. |
| Compatible API | /models and /chat/completions SSE for planning/review. No filesystem coding tools. |
Provider installation, authentication, and actual inference are separate states. See setup and capability evidence and Hermes connector research.
flowchart LR
A[Describe the change] --> B[Review the plan]
B --> C[Approve execution]
C --> D[Isolated coding leads]
D --> E[Helper reviews]
E --> F[Lead integrates findings]
F --> G[Verify and review]
G --> H[Approve final patch]
Coding teams use separate managed Codex sessions on the task's chosen coding configuration. Helpers review in read-only mode; the lead writes and integrates. Findings are bounded, redacted, and treated as untrusted evidence. Coding-team details.
npm test # deterministic tests; no external model calls
npm run test:e2e # build + real Electron UI/persistence checks
npm run format:check
npm run test:widget # floating widget and native lifecycle
npm run test:live-ui # live graph, text direction, and Ava
npm run test:provider # actual Codex workflow in a disposable repository
node scripts/package-smoke.mjs # run after packagingThe v0.1.9 implementation passed 84 automated tests, desktop acceptance, actual Codex team execution, and packaged launch/persistence checks. The real-provider smoke uses your current account and consumes model usage; it automatically approves only its narrowly defined disposable fixture. Verification evidence.
npm run analyze:graph uses an installed Graphify CLI for code-only AST extraction. npm run benchmark records a disposable context/streaming benchmark and CPU profile. Scoped source excerpts, related imports, bounded handoffs, and batched stream persistence keep context and UI work bounded. Token estimates are not billing measurements. Performance notes.
On macOS, application data lives in ~/Library/Application Support/aether-code/. CLI credentials stay in the provider's credential store; Aether does not copy OAuth tokens. An optional compatible API key uses Electron safeStorage and is not returned to the renderer. No .env file is required for CLI use.
The Electron renderer has no Node access and uses validated IPC. Provider subprocesses have bounded output, timeouts, cancellation, a restricted environment, and their configured sandbox. Approved verification commands execute repository code on your host outside the Codex sandbox. Worktrees isolate changes, and scope checks detect violations after execution; they are not an operating-system sandbox.
Use trusted projects and inspect approved commands. The metadata database is unencrypted, and retained worktrees contain source and execution evidence. See security boundaries and remaining gaps.
- Architecture and module boundaries
- Automatic builds and releases
- Contributing · Implementation checklist
- Nexload prompt core
- Coding teams · Live graph
- Ava companion · Desktop widget
- Provider setup · Development environment
This is a developer preview. Remaining work includes hardened verification isolation, additional verified coding providers, dependency provisioning, stronger budget enforcement, worktree retention controls, and signed distribution with broader cross-platform runtime validation.
MIT. Third-party fonts and vendored components retain their own licenses; see notices. Aether's compact prompt foundation adapts the pinned Nexload reasoning system, Ava uses an adapted Bloub animation core, and live source exploration uses Graphify.