A voice coding assistant you can actually talk to

“Talk to AIDEN” is a voice interface wired into the agentic IDE. When you speak, AIDEN either acts instantly inside the app, open a panel, split a view, read something back, or hands the request to a background agent that actually writes the code. You already delegate features to agents from the board; voice lets you kick them off and steer them without touching the keyboard.

You open it with Cmd/Ctrl+Shift+V, hold Space to talk, and release to send. AIDEN replies in a natural, low-latency voice over WebRTC, powered by OpenAI’s Realtime API (model gpt-realtime-2, with selectable voices like marin, cedar, and alloy). Your words and AIDEN’s reply stream as live captions, so the current session always has an on-screen transcript. Change your mind mid-answer? Hold Space to barge in and start a new turn, or hit Esc to hard-stop a response.

Under the hood, AIDEN decides whether your request is a quick in-app action or real work that needs an agent. Here is the full path from spoken words to a delegated result.

1. 1

   Push-to-talk

   Toggle voice mode with Cmd/Ctrl+Shift+V, hold Space, speak, release to send. No wake word, no always-on listening, the mic is only live while you hold the key. Your speech is transcribed as streaming captions.
2. 2

   Fast tools or delegate

   For lightweight things, navigating the app, splitting or opening panels, a quick read, the voice model calls one of about three dozen curated fast tools and the action happens immediately. Anything heavier (coding, research, file edits, browsing, automation) goes to delegate_task instead.
3. 3

   Background agent spawns

   delegate_task spawns a background agent and routes the work to the best worker: Codex CLI for coding, Claude CLI for general tasks, with the OpenAI API as a fallback. Voice runs one delegated agent at a time, separate from the parallel agents on your board.
4. 4

   Approve risky actions

   Destructive or risky actions, deleting, sending, deploying, calendar or account changes, are auto-flagged and paused. Approve and Decline buttons appear in the voice HUD, and nothing runs until you say yes.
5. 5

   Monitor and hear the result

   A top-bar HUD shows the running agent count and narrates the current step and tool, moving through queued, running, and completed or failed. When the task finishes, AIDEN speaks the result summary aloud.

The delegated agent runs on the same local CLIs as the rest of AIDEN, the orchestration layer described in Claude Code orchestration. Voice is the single-slot entry point; for many agents shipping in parallel, see the multi-agent coding workflow.

The voice orchestrator is built around a small set of verified capabilities. This is exactly what it does today.

Typing-only assistants are excellent, and AIDEN is one by default, everything works from the keyboard and the board. Voice adds a second input mode for the moments talking is faster: kicking off a task while your hands are busy, steering an agent mid-run, or thinking out loud about what to build next.

Voice does not replace typing, it removes friction at the exact moments the keyboard gets in the way. Delegated features still land as cards on the AI kanban board, you still review every diff, and your code never leaves your machine.