Aether Agent

Coding and account agents in your terminal.

It reads your repository, makes the change, runs the checks you name, and shows you the exit code. Hosted models or your own local Ollama. The 0.4.0 source candidate also brings your account agents and Online chats into the terminal, with guided ATS setup.

Quickstart · Models · Commands · Aether Code · Safety · Patch notes

_{Install → launch → /help → /models. That's the whole first run.}

Requires 0.3.2 or newer. Check what you have with aether --version.

Quickstart

Run these from inside the repository you want Aether to work on:

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

That last line is the whole idea: one task, one way to prove it worked. Aether makes the edit, your machine runs npm test, and the real exit code decides whether the run is verified. No exit code, no claim.

Prefer Python? pipx install aether-agent installs the same CLI and forwards every command to it, so aether-agent code "..." and aether code "..." do the same work. See packages/pypi-cli.

Account agents and ATS — 0.4.0 source candidate

ATS is the trading adapter for account agents.

After signing in, use aether agent list to see the same managed agents as Aether Online. aether agent chat opens the one-column picker; select an agent to read and send messages in its existing Online conversation.

aether agent list
aether agent create ATS Atlas
aether agent configure <agent-id> purpose "Review my trading strategies"
aether agent chat <agent-id>

ATS setup asks for a local memory folder, a memory size, a strategy folder and data provider/symbol settings. Local resources belong to the verified account and agent; rotating a token preserves that identity. Switching accounts while chat is open closes its local resources and requires reopening the conversation. Before the first setup for a policy version, the terminal presents the ATS Autonomous Trading Acceptable Use, Risk and Data Policy: choose 1 to accept or 2 to reject. Rejection exits before creating an agent or configuring storage, strategies, datafeeds, browsers, plugins or MCP. The local, pseudonymous consent receipt records the policy digest but grants no broker or trading authority. The packaged adapters verify storage and report native Nano compiler results. Inside ATS chat, type /ats status, /ats strategies or /ats data to inspect the workspace. Shift-Tab cycles the local ATS permission preference. Model availability, projects, APR memory and UVT admission remain enforced by Cloud.

ATS chat opens the configured browser view after memory verification. For any managed agent, /browser setup [URL] configures the connection and /browser open opens it. Use /browser status, /browser refresh, /browser stop or /browser retry to manage the view; /ats browser is an alias. Status updates preserve your draft and cursor. LIVE requires a fresh validated screenshot; an open window alone is not proof of observation. Retry starts a new bounded session after releasing the old one. Cleanup receipts survive restarts. If a creation response was lost, an idle runtime cannot prove that request finished; the terminal retains the receipt and blocks replacement until cleanup is known.

The local API defaults to http://127.0.0.1:8092. Run npx aether-browser@0.2.2 doctor to check the separately installed runtime. Remote APIs use HTTPS and environment credentials. The native noVNC viewer stays on the browser host and must never be tunneled or published. The bundled read-only visual skill exposes verified images for an admitted host; Cloud DM does not yet receive those images or control this browser.

These commands require the matching Cloud terminal adapter. ATS setup also requires its Python engine and a separately running Agent Browser runtime; missing services are reported explicitly. This candidate prepares and observes an ATS workspace. It does not yet provide model-controlled broker actions or automatic live orders. See the candidate packet for the qualification still required before publication.

Pick where the model runs

Both routes keep your files, your permissions, and your verification on your machine. The only thing that moves is where the model thinks.

Hosted — sign in and go

aether auth login
aether models

Your task and the context you hand over go to the Aether API. Repository tools and checks still run in your checkout. If the service can't honour that local-authority contract, the run stops rather than quietly moving your tools to a server.

Local — no account needed

Install Ollama, start it, then:

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"

Once Ollama and the model are downloaded, inference can stay on the machine. The default endpoint is loopback; if OLLAMA_HOST points elsewhere, prompts go there instead. Network tools stay separate and permissioned either way.

Model catalogue

What your account can actually reach is whatever aether models prints while you are signed in. The snapshot below is a dated reference, published so the list is readable without signing in first.

A dated, sanitized offline fallback snapshot is available as HTML, JSON, and Markdown. It was generated at 2026-08-23T00:00:00.000Z from Cloud public projection model-catalogue-v1 with verified digest sha256:80ba3ba1144d301e2cca407ceced74cb2b371f1da6e3982b87305ff12a3d4712. Listed availability is not an account entitlement; use aether models while signed in.

Local Ollama is independent of all of it — you get whatever you have installed at your configured endpoint.

What you get

• One terminal workflow. Task, diff, tests, and review in the same place you already work.
• Proof, not claims. --test-cmd ties "done" to a command and an exit code. Change the repo afterwards and that evidence goes stale on purpose.
• Repo-aware tools. File, search, shell, Git, session, and review tools scoped to the workspace you chose — not your whole machine.
• Work you can pick back up. Sessions are project-scoped: list them, resume one, or hand it off as a redacted bundle.
• MCP built in. Inspect, diagnose, and repair configured MCP servers without stepping around tool permissions.

Commands

Command	What it does
aether auth login	Sign in for hosted models.
aether agent [task]	Run the coding agent, or open its REPL.
aether agent --local [task]	Same, through your Ollama endpoint.
aether models	Show the hosted models your account can see.
aether local doctor|models|use|pull	Diagnose and manage local Ollama.
aether sessions	Inspect and continue project-scoped sessions.
aether review	See the changes and the current verification evidence.
aether mcp	List, diagnose, or repair MCP servers.
aether doctor	Check the environment.
aether ship	Preview and approve a branch and pull request.

aether help <command> has the details, or read the generated command reference for every flag, slash command, environment variable, and exit code.

Aether Code on the web

Aether Code is the browser surface for the same Aether account — one of three apps on the portal, alongside Web Chat and Design Lab. Sign in once, then pick where you want to work that day.

Aether Code and Aether Agent are deliberately separate products. The CLI is standalone and open source: on the local route it needs no Aether account at all. Coding workspace sessions stay on their host; the managed-agent workflow above shares account agents and their Online DM conversations.

Coming next: live session viewing

Remote viewing — /rc — is the bridge between the terminal and the browser, and it is being integrated now. The host lives in PR #108 and is not part of 0.3.x, so nothing below is something you can run yet. What it will do:

• Starting a session prints a link and a QR code.
• Your phone or browser watches the run. It never gets tool authority.
• One command shows what is exposed; another revokes it.
• Outbound TLS only — no inbound listener on your machine.
• Events are allowlisted and redacted: no environment variables, credentials, cookies, private memory, raw file contents, absolute paths, or unredacted shell history.
• If the broker drops, your local session carries on regardless.

Those are release requirements, not goals — which is why the work is still open rather than shipped.

What stays on your machine

• Files stay in the workspace. File tools are confined to the workspace you selected, and writes, shell, Git, network, and publishing each pass a host-side permission gate.
• Hosted runs send only what you hand over. Your task and context go to the Aether API; repository tools and checks run in your checkout, and the local-authority route fails closed rather than quietly degrading.
• Local runs stay local. Ollama prompts go to your configured endpoint, which is loopback by default.
• Secrets live outside the repository. So do session records. Portable handoffs drop transcripts, file contents, shell commands, and absolute paths — still give one a read before you share it.
• Publishing is never implicit. aether ship prints its branch, commit, destination, and pull-request plan before it acts, and --yes on its own is not publication authority.

SECURITY.md has the supported versions, the full boundary, and the private path for reporting a vulnerability.

Versions and patch notes

The repository and the published package are versioned independently.

Install	Version	What it is
npm latest		The published package; the badge resolves the live dist-tag.
PyPI aether-agent		A launcher that installs and runs the npm CLI. It follows npm latest unless you pin one.
main source build	0.4.0	Source candidate for shared agents and ATS setup; publication and live-service qualification are recorded separately in the operator packet.

• Release notes — one entry per release, in plain language.
• Release log — dated candidate, publication, and documentation events, plus operator packets that freeze prerelease qualification and link the authoritative tag and registry evidence.
• Releases and tags — every published version.

Build from source

Node.js 24 or newer:

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

Development

npm ci --ignore-scripts
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 the reviewed aether-ats-skills source package with exact aether-browser and aether-context dependencies. TypeScript and Node types remain development-only. Read CONTRIBUTING.md before opening a pull request, and see the architecture and protocol docs or production operations if you are going deeper.

Help and license

Bugs and feature requests go to GitHub issues. Security reports use the private path in SECURITY.md.

The Agent and bundled ATS adapter code are Apache-2.0. The paid ATS engine is a separate prerequisite. The license covers the code, not the Aether name or the hosted service (LICENSE · NOTICE.md).