Voice Copilot

What it is

Voice Copilot runs alongside CLI coding agents like Claude Code and Codex. A cheap model narrates what the agent is doing by turning event streams into spoken updates, while a stronger model supervises at checkpoints, warning you aloud when the agent drifts and pausing it if needed. Both reuse the CLI you already launched, without requiring extra API keys.

Instead of babysitting the terminal or missing important mistakes, Voice Copilot keeps you aware of what the agent understood, what it changed, and what risks it took. The Supervisor catches problems while they can still be stopped, and you decide whether to resume or redirect the agent based on what you hear and see in the trace.

Questions from the bench

What is Voice Copilot not?

Voice Copilot is not a voice keyboard for dictation, not raw terminal text-to-speech that reads every line, and not another autonomous agent. The Supervisor never writes code - it watches the agent that does and hands the decision to you.

Which agents does Voice Copilot work with?

Voice Copilot runs beside Claude Code, Codex, and other CLI agents.

How does the Supervisor work?

The Supervisor uses a strong model to review the agent's work at checkpoints rather than every event, spending tokens only where they count. It can issue three verdicts: OK is silent, WARN is spoken, and STOP is spoken and pauses the agent with Supervisor+.

What are the checkpoints that trigger supervision review?

Checkpoints include the end of the turn, every N tool calls, and failures that repeat. Sub-agents finishing and the CLI's own side requests do not count.