Wicked Studio 0.14 lets you steer coding-agent jobs by voice

Wicked Studio’s new release moves coding-agent supervision into a voice conversation. Version 0.14.0, published November 2 at 08:04 UTC, adds a workflow for selecting a running job by speaking, inspecting its live state and telling it what to do next. The project says each action is confirmed on screen before it is sent.

Studio is the browser interface for wicked-crew, which runs existing coding-agent CLIs such as Claude Code, Codex, GitHub Copilot, OpenCode and Pi. This is an operator interface update, not a new model or a demonstrated improvement in coding accuracy.

Find the job, then give the instruction

The release describes spoken job selection using the job in view, the only running job or a fuzzy match against recent job names. Ambiguous matches produce a choice rather than silently picking one. A live tile brings together the job’s phase, checks, newest activity and any waiting decision.

Once a job is selected, speech can propose a steering note, a working-document instruction, a stop, a higher rigor setting or an answer to a waiting decision. The notes say explicit phrases map directly to actions; unmatched wording can use a system-chat classification. Both routes still lead to confirmation. The thread marks supported voice-originated notes as “You (by voice).”

This extends the earlier chat-first job page with a concrete way to supervise work while away from the keyboard. It does not mean a spoken suggestion automatically authorizes a job to act.

Local transcription still needs the right backend

The documented speech path sends audio to Crew’s local transcription endpoint, which invokes whisper.cpp on the same machine. Crew’s documentation says the engine must be installed separately and its speech model downloads only with consent. Replies use browser speech synthesis, with captions available. Local transcription should not be confused with local coding inference: the selected agents still use the operator’s own provider accounts and allowances.

General setup uses Node.js 22 or newer, npm 10 or newer, a container runtime and wicked-garden 12.40.0 or newer for governed workers. Studio normally ships inside Crew through npx wicked-crew serve; its README also documents building the frontend separately against a daemon.

There is a material version caveat. At review time, the latest Crew GitHub release and npm package were 0.16.11, dated October 29, and that package declared Studio ^0.13.0. Those facts do not establish that the normal published Crew bundle already includes Studio 0.14. Voice also requires a daemon exposing the corresponding endpoints; the frontend release does not provide a verified minimum backend version for every new action.

What has been checked

AITrending read the English release, tagged setup documentation and voice-flow test source without installing the application. The tests use a fake microphone and stub speech-to-text, so their presence is not evidence of real-world recognition accuracy or end-to-end performance. The supported news is a released frontend with a documented voice-supervision workflow, with matching-backend availability still something to verify before deployment.