@aetheros/voice

The Aevron voice console: a local page you talk to by your project's assistant name. It works in any project; the name and colours come from .aether/brand.json (for example "Nova" in "Nova OS").

aether voice                  # http://127.0.0.1:4747/ — open in Chrome or Edge for voice
aether voice --ai claude      # also answer free questions (ANTHROPIC_API_KEY)
aether voice --ai claude --tier cheap         # quicker spoken answers from the catalog's cheap model
aether voice --ai mistral --model <model id>  # any provider: claude, openai, google, xai, mistral, local
aether voice --agents         # your workspace agents answer, with their tools, memory, budgets and Review
aether voice --agents --agent reviewer       # a different default agent
aether voice --pack packs/app # which flow pack "run full scan" uses (default: .aether/voice.json or the first packs/*)

Say "Nova, run full scan" (with your own assistant name), click a command, or type.

Workspace agents (--agents):

• Questions go to the first workspace agent (or --agent <id>), in one saved conversation per agent that you can open later in aether workspace. Say "ask Reviewer to check the README" to ask another agent.
• The agent's tools, memory and daily budget apply as in the workspace.
• When an agent wants to change something, the console reads the request out at once: "Nova wants to use owner.add_item: … Say confirm to approve it, or deny." The agent's final answer is spoken when the request is decided.
• "what needs review" lists waiting requests; "approve" reads the first one out; "confirm" approves the one just read; "deny" refuses it. Nothing is approved without being read out first.
• The console decides its own requests; they don't appear in the workspace page's Review tab. Requests left undecided are refused after 10 minutes.

Talking with it:

• Animated HUD: a ring around the orb follows your microphone while it listens, turns while it thinks and pulses while it speaks. With reduced motion it stays still.
• Continuous conversation: after "Nova", keep talking without the name for 20 seconds. It remembers the last few turns, so "and after that?" works. Say "thanks", "that's all" or "stop listening" to finish.
• Interruption: start talking while it speaks and it stops at once, cancels the answer still on its way (also on the server), and listens. "Stop" or "quiet" only silences it.
• Streaming: with a provider that streams, answers are spoken sentence by sentence as they arrive.
• Headphones help: while it speaks it ignores phrases that are mostly its own words, but speakers can still confuse the browser's recognition.

| You say | It does | |---|---| | run full scan | Runs the flow pack with aether run in the background, then says how many flows passed or failed and which ones failed. Reports go to .aether/runs/. | | what's waiting on me | Reads "Waiting on the owner" in HANDOVER.md plus open owner-list items. | | what's building | Agents running now, plus open pull requests and their checks (with GITHUB_TOKEN). | | what's next | "Up next" from the handover. | | any ideas | Ideas to decide, and approved ideas not set up yet. | | anything else | Answered in a few spoken sentences by the AI through Aevron (with --ai), using only a short project summary. |

Safety:

• It only listens on localhost.
• Commands need the page's per-run token and a localhost Host header, and the page runs one script allowed by a nonce.
• The microphone stays off until you press Start listening.
• While it's speaking it ignores phrases that are mostly its own words, so it doesn't answer itself.
• The conversation history stays in the page and is sent with each question (last 12 turns); nothing is saved.
• It only reads the project, apart from flow reports.

Limitations:

• Voice uses the browser's speech recognition, which works best in Chrome and Edge. Typing works everywhere.
• Free questions don't take actions; only the listed commands do.
• One project per console.