Aether Agent

Build, test, and review from your terminal.

An open-source coding agent that reads your repository, makes changes, and runs your checks. Choose hosted models or local Ollama. Keep working in the same terminal.

Quickstart · v4.21.0 update · Features · Models · Terminal · Docs · Contribute

_{Install → launch → /help → /models}

What you can do

Workflow	What you get
Code and verify	Repository search, file edits, Git tools, and checks with real exit codes.
Chat and use your shell	Run !commands between messages; open an interactive terminal on Linux. Terminal guide ↓
Choose your model	Hosted account models or local Ollama with no account required. Model setup ↓
Continue your work	Resume project sessions, review changes, and export redacted handoffs. Commands ↓
Connect tools	MCP servers, agent skills, and local development previews. Command reference
Use account agents	Shared Online conversations and guided ATS workspace setup. Account agents ↓

Quickstart

Node.js 24+ required. Run inside the repository you want to work on:

npm install -g aether-agents@latest --ignore-scripts
aether auth login
aether

Type /help for commands or /models to choose a model. For a coding task with a specific check:

aether agent --test-cmd "npm test" "fix the failing test"

For a multiline task, save the specification in task.md and run aether agent --prompt-file task.md. You can also pipe UTF-8 text to aether agent --prompt-file - (for example, cat task.md | aether agent --prompt-file -). The whole input is one literal coding task; see prompt input rules for the size limit and flag conflicts.

To inspect the current checkout and receive a plan without changing it, run aether agent --planning "outline the migration", or use /plan <topic> in a local interactive session. Planning permits only file reading, directory listing, and repository search; it skips worktree creation and verification. Save a plan or execute a phase later through the explicit goal controls. The cloud chat route refuses /plan because its tools execute on the server.

A completed check records its exit code. Changing the repository makes that verification stale until you run it again.

The npm version badge shows the published CLI version. Run aether --version to check your installation. For the latest published build, use npm install -g aether-agents@latest --ignore-scripts.

Prefer Python? pipx install aether-agent installs a launcher for the same CLI. Python setup.

What's new in v4.21.0

• Find and enter commands faster. The idle composer describes slash commands, searches earlier prompts with Ctrl+R, and supports multiline editing with undo and yank.
• Keep task context under your control. Pin bounded files for a coding turn, inspect what reached the model, and run a trusted skill for one local turn.
• Use safer task inputs. aether agent --prompt-file accepts one literal specification, while planning runs with read-only host tools. Tool denials carry the operator's feedback into the failed result.

See the full v4.21.0 release notes for fixes, install options, and qualification limits.

Choose your model

Hosted	Local Ollama
Sign in with aether auth login.	No Aether account required.
aether models shows your account's available models.	Use a model installed at your Ollama endpoint.
Coding tasks send prompts and selected context to Aether; tools and checks run in your checkout through aether agent.	Inference stays at your configured endpoint, which defaults to loopback.

To use Ollama, install and start it, then run:

aether setup --local
aether local pull qwen2.5-coder:7b --yes
aether local use qwen2.5-coder:7b --yes
aether agent --local --test-cmd "npm test" "fix the failing test"

If you set OLLAMA_HOST to a remote endpoint, prompts go there. Network tools have their own permissions.

Browse the offline model catalogue

A dated, sanitized offline fallback snapshot is available as HTML, JSON, and Markdown. It was generated at 2026-09-22T01:47:56.169Z from Cloud public projection model-catalogue-v1 with verified digest sha256:f5f516625932d8932bfca221aa5dbf3eb1d7b415eba8dabe298da24b984c64f7. Listed availability is not an account entitlement; use aether models while signed in.

Chat and terminal

In the local coding console, switch between conversation and your own commands:

Explain the failing test
!npm test
/shell-result
/shell-result lines
/shell-result drop 8-10
/shell-result send

!commands run locally with zero model API calls. Output streams into the console with an exit code. /shell-result stages and previews the exact bounded attachment, including the command, captured directory, exit status, and any omitted bytes. Edit it with drop, replace, mask, or redact, then choose /shell-result send or /shell-result cancel. Redaction helps identify common secrets but does not guarantee their removal. A later shell command cannot change the staged attachment. For pipes and JSON sessions, /shell-result send is the explicit one-step send form.

Platform	Local shell	Interactive programs
Linux	Persistent Bash: keeps cwd, exports, and functions.	/terminal python3 starts a terminal with input and resize support. Requires Python 3 and TTY input/output.
macOS	Persistent Bash: keeps cwd, exports, and functions.	Use an external terminal.
Windows	One-shot cmd.exe by default. /shell-profile use powershell explicitly starts a persistent PowerShell session for cwd, environment values, and functions.	Use an external terminal.

On Windows, run /shell-profile list to see the installed executable and version before choosing PowerShell. /shell-profile status shows the active profile and cwd; /shell-profile use cmd returns to the compatibility profile. Switching or /shell-reset discards shell state without replaying commands. PowerShell commands keep their own quoting and syntax; Bash commands are not translated.

Shell: Ctrl+C cancels the command; /shell-reset starts fresh after cancellation or failure. Commands wait for an active model/tool turn to finish. Linux terminal: Ctrl+] returns to chat; /terminal-attach reconnects; /terminal-stop ends it. Local tools pause while that terminal is running, including when detached.

Shell and terminal output stays out of saved chat history and automatic hosted prompts. Account-agent DMs use a separate console. Full shell and Linux guide.

In the raw terminal composer, Ctrl+J adds a newline, Enter submits, Ctrl+_ undoes recent edits, and Ctrl+Y restores killed text. The single-row input shows newlines as ⏎ with a line count. If your terminal sends LF for Enter, set aether config set lfSubmits true to keep LF as submit. Composer details.

Ctrl+R searches this workspace's submitted prompts newest first. Type to narrow matches, repeat Ctrl+R for older ones, Enter to insert a match for editing, or Escape to restore your draft. AETHER_NO_HISTORY=1 disables prompt history.

If you decline a model-requested tool call, you can add one short instruction for that denial, such as Use the offline unit suite. The tool stays blocked; the instruction is returned once with its failed result. Enter skips the note, and Escape cancels it. Approval details.

Type / at an idle raw terminal prompt to browse described commands. Use arrows or Tab to choose, Enter to insert editable command text, then Enter again to run it; Escape restores your earlier draft.

On a local model, /skill user/fix-ci Investigate the failing unit test uses that trusted skill for one idle task, then restores your normal skill defaults. The task text stays literal; the run header shows the resolved skill digest and host tool limits. Busy, queued, line-input, --no-skills, and server-executed cloud chat invocations are refused. Skill details.

Use /pin src/guide.md to attach that file's complete, bounded UTF-8 content to each admitted coding turn. The file is read from the execution checkout, so a worktree turn uses its own copy. /context shows what the last turn actually included, with digests and omissions but no file bodies; /context next <task> previews a draft, and /context content src/guide.md explicitly previews local admitted content. /drop src/guide.md stops automatic inclusion on later turns. Server-executed cloud chat reports pin delivery as unsupported when the local host cannot inspect it. Limits and details.

Everyday commands

Goal	Command
Run a coding task	aether agent "your task"
Continue a project session	aether sessions
Review changes and verification	aether review
Connect or diagnose tools	aether mcp, aether skills
Manage a local dev preview	aether preview
Check your environment	aether doctor, aether pc map, aether pc doctor
Configure the CLI	aether settings
Preview and approve a branch and PR	aether ship

Use aether help <command> or the complete command reference for flags and slash commands. PC diagnostics and approved browser helpers have platform-specific limits; see the PC guide.

Account agents and ATS

Use the same managed agents and Online conversations from your terminal:

aether agent list
aether agent chat

aether agent list and show report fresh registry, DM, and model/UVT readiness for the signed-in account. aether doctor --live checks the same read-only Cloud contract; plain aether doctor leaves account readiness unverified. A saved DM is reported as admitted only when Cloud confirms message admission. In aether agent chat, in-chat help lists controls for the selected agent and can refresh the shared conversation. The header shows Cloud DM sync separately from ATS local setup. Transcript messages show their time and any reported admission state; a saved message is not an admitted run.

ATS is the trading adapter for account agents. Create a workspace with aether agent create ATS Atlas; guided setup covers memory, strategies, and data settings.

Capability	Current status
Shared account agents and Online chats	Requires the matching Cloud adapter and account admission.
ATS workspace and strategy preparation	Local memory setup, native Nano compiler checks, and data-provider configuration.
Browser observation	Read-only view; requires a separately running Agent Browser runtime.
Market-data execution and orders	Production runtime wiring and managed-agent order execution are unavailable.

ATS requires the separate Python engine and policy consent. This build does not yet provide model-controlled broker actions or automatic live orders. Setup and observation do not grant trading authority.

Account agents and ATS guide · ATS policy · Release qualification

Web and remote viewing

Aether Code is the browser coding app alongside Web Chat and Design Lab. It uses the same Aether account; the CLI also works independently with local Ollama. Coding sessions stay on their host, while managed agents share their Online conversations.

Remote viewing (aether rc) is included in v4.20.0. A live Cloud viewer journey (Windows and Linux hosts, phone viewer) was recorded on 2026-10-06. aether rc start requests a separate owner-scoped RC identity and does not require operator device enrollment; an ordinary-account deployed journey still needs qualification. It is designed to let a browser or phone watch a redacted terminal run through a link or QR code. The viewer has observation access only. Remote viewing status and controls.

Privacy and control

• You choose the workspace. File tools enforce its boundaries. Shell commands run with your user's permissions; the workspace directory is not an OS sandbox.
• You control execution. Model tools pass host permission gates. Running your own shell command does not authorize future model actions.
• You choose what to share. Hosted coding sends task context to Aether. Local Ollama uses your configured endpoint. Shell output requires explicit sharing; credentials and session records live outside the repository.
• You review before publishing. aether ship shows the branch, commit, destination, and PR plan before approval. --yes alone cannot authorize publication.

Security policy and private reporting. Review redacted session exports before sharing them.

Documentation

Start here	Reference
Commands and configuration	Command reference
Shell and Linux terminal	Local shell guide
Account agents and ATS	Setup, browser controls, and execution status
PC and browser capabilities	PC guide
Remote viewing	Status and controls
Project memory	Memory guide
Architecture and operations	Docs directory · Production operations

Build and contribute

Build from source with Node.js 24+:

git clone https://github.com/AetherAI3/aether-agent.git
cd aether-agent
npm ci --ignore-scripts
npm run build
npm link

Start with CONTRIBUTING.md for the repository map, development workflow, and required checks. Bug reports and feature requests go to GitHub issues.

Development checks

npm run typecheck
npm test
npm run smoke
npm run verify:production
npm run docs:check
npm run release:truth
npm pack --dry-run

The runtime bundles reviewed ATS adapter code and pinned browser/context dependencies. See the contributing guide before changing that dependency graph.

Versions

Install	Version	What it is
npm latest		Published CLI; the badge resolves the live dist-tag.
PyPI aether-agent		Python launcher for npm latest, unless pinned.
main source build	4.21.0	See the v4.21.0 notes and operator packet.

Release notes · Release log · Releases and tags

License

Apache-2.0 for the Agent and bundled ATS adapter. The paid ATS engine is a separate prerequisite. The license covers the code; the Aether name and hosted service have separate terms. LICENSE · NOTICE.md.