Voice production is easy to automate badly. An agent can generate a folder of files in minutes, but the useful outcome is a local project with approved scripts, chosen voices, and exports that fit where the audio will be published.
Vois gives an agent a controlled path to that outcome. The desktop app remains the local studio, and its CLI lets an agent create projects, prepare scripts, generate approved audio, and export with the chosen profile. Your scripts and audio stay on the same machine.
The difference is important: do not give an agent a pile of guessed terminal commands. Give it the Vois skills, ask it to plan the workflow, and keep voice, script, and export approval with the person responsible for the result.
What the CLI is for
The Vois CLI ships with the desktop app. It communicates with the running app through a local connection, so the app performs generation, project management, and audio processing while the agent handles repeatable instructions around it.
The agent can work across the production workflow:
| Workflow area | What the agent can prepare or run after approval |
|---|---|
| Speech generation | Draft clips, preview voices, and batch approved script sections |
| Voice management | Search the library and organize authorized clones |
| Projects and scripts | Create local projects, import text, and structure speakers |
| Production | Manage timeline data and choose a mastering profile |
| Export | Produce WAV, MP3, FLAC, or AAC files for the intended destination |
| Pronunciation | Propose and maintain project-specific pronunciation entries |
The CLI has 15 command groups and 60+ commands. That breadth is why the skill files matter. An agent that has read the current reference can inspect the available capability instead of inventing syntax or assuming a cloud API is available.
Give the agent the right instructions
Vois hosts two agent-facing skills:
- CLI reference, covering installation, local connection requirements, command groups, output handling, errors, and production workflows.
- Dialogue-writing guide, covering natural spoken scripts, speaker tags, pacing, and dialogue formats.
Prompt your agent: "Read the Vois CLI and dialogue-writing skills. Inspect the running local Vois app and propose a production plan for my project. Include project structure, speaker assignments, voice candidates, a preview plan, and an export profile. Do not modify my shell configuration, create voices, generate audio, or export files until I approve the plan."
Expected deliverable: a concise plan with the local project name, script outline, proposed voice IDs, one short preview line per speaker, and an export target.
Review and approve:
- Confirm the app is running and the agent has found the bundled CLI without changing your system configuration.
- Review the script, approved voice sources, and proposed destination before any generation.
- Approve short previews, then listen to each one in Vois and choose the final voice assignments.
- Approve batch generation only after the previews are right, then review the mastered export before it is distributed.
This keeps the agent useful without handing it editorial or publishing authority.
A practical example: a podcast series
Suppose you want five episodes about startup fundraising, each with a single host and a consistent delivery. The agent first reads the dialogue-writing guide and proposes spoken scripts with natural paragraph breaks. It then reads the CLI reference and lays out one local project, five scripts, a host voice candidate, preview clips, and an Apple Podcasts export profile.
You review the scripts and approve one short preview for the host. Only after that preview sounds right does the agent create the project content and generate the approved episodes. The resulting clips stay organized in the Vois project, where you can inspect the timeline, listen for pacing, and approve the final mastered exports.
That is the useful automation boundary. The agent handles repetition. You decide what is said, who it sounds like, and what gets released.
Set it up without copying terminal commands
Install Vois and open the desktop app. The included CLI is a local remote control for that running app, not a separate hosted speech service. If your agent needs terminal access, let the current CLI skill guide it through locating the bundled binary and checking the local connection.
Prompt your agent: "Read the current Vois CLI skill and verify whether the bundled CLI can communicate with the open Vois desktop app. Report the installation location and connection result. Do not create aliases, change PATH, run privileged commands, or generate audio unless I explicitly approve those actions."
Expected deliverable: a read-only setup report that says whether the app is reachable, what the agent needs next, and whether any optional local configuration change would be required.
Review and approve:
- Check that the report describes a local connection to the open Vois app.
- If optional shell configuration is proposed, review the exact change before allowing it.
- Ask the agent to make one disposable preview only after you approve the voice and text.
- Listen to that preview in your normal player before using the CLI in a larger workflow.
What still belongs in the app
The CLI is strongest at repeatable production: batches, projects, scripts, speaker assignments, and exports. Some decisions are better made with the visual studio open. Arrange clips by ear, inspect waveforms, trim timing, and make final creative calls on the timeline.
That split is a feature, not a limitation. A coding agent can remove repetitive setup, while Vois keeps the production review in the place where you can hear what the audience will hear.
Keep a small approval record with the project: the final script, selected voice IDs, preview decision, mastering profile, and export destination. It gives the next batch a clear starting point and prevents an agent from treating a prior experiment as a publishing decision.
Start with a small approved job
Do not start with a fifty-episode batch. Give the agent one paragraph, one approved library voice, and one file destination. Review the sample. If it is right, reuse the plan for the next project. If it is not, adjust the script or voice before scale turns a small mismatch into dozens of files.
Read the CLI automation feature guide, then Get started.
The Vois Team