Prompt dictation
It is not a voice keyboard. You speak when you need to steer the agent.
Open-source alpha 0.1.0 for LLM coding CLIs
Listen to Control
A cheap model narrates your coding agent. A strong one supervises it. You step in the moment it matters.
Voice Copilot runs beside Claude Code, Codex, and other CLI agents. The narrator turns the event stream into short spoken updates. The Supervisor reviews the work at checkpoints, warns you out loud when the agent drifts, and can pause it until you weigh in. Both reuse the CLI you launched, with no extra keys.
Narration active
/
Supervisor watching
/
agent connected
Why it exists
Modern agents can work for minutes. Your attention should not be trapped there.
You either babysit the terminal or look away and find out about a bad turn after it happened. Voice Copilot keeps you aware of what the agent understood, what it changed, what risk it took, and puts a second, stronger model on watch so the bad turn gets caught while it can still be stopped.
How it works
Two models, opposite budgets.
$ vc codex "fix the two failing tests"
agent reading invoicer/money.py and the failing pytest output
tool exec python -m pytest -q
edit invoicer/invoice.py
Two failures located: the discount is rounded too early in money.py, and VAT is charged on the pre-discount subtotal. Two small edits are going in, then pytest again.
OK 4 tool calls since the last review
OK turn ended, tests green
WARN tests were not run after the last edit
STOP editing files outside the task - agent paused
I paused the agent. It started editing hotkeys.py, which has nothing to do with the failing tests. Resume, or tell it to stay in invoicer/.
The Supervisor
A stronger model on watch, spending tokens only where they count.
Reviewing every event with a capable model would be slow and expensive. Reviewing a compressed transcript at a few checkpoints costs a fraction of one agent turn, and that is exactly where a strong model earns its keep: noticing that the worker is editing the wrong files, looping, running something destructive, or calling a job done that is not.
- Checkpoints, not a ticker. The end of the turn, every N tool calls, a failure that repeats. Sub-agents finishing and the CLI's own side requests do not count.
- Three words. OK is silent. WARN is spoken. STOP is spoken and, with Supervisor+, pauses the agent until you resume it from the panel or the hotkey.
- Your CLI, your models. One checkbox picks the weakest model for the narrator and the strongest for the Supervisor from the CLI's own catalog. No API keys, no second login.
- Per CLI. Turn it on for Codex only, give Claude Code its own model, or leave it off where you trust the agent.
USER fix the two failing tests
TOOL exec: python -m pytest -q
NARRATION Two failures located; two small edits going in.
SUPERVISOR OK - 4 tool calls since the last review
FILE src/voice_copilot/hotkeys.py
SUPERVISOR STOP - agent paused: editing files outside the task
PAUSED agent paused (supervisor)
Demo
See the loop in under a minute.
The demo is published on YouTube, so visitors can understand the product before reading docs.
Positioning
What Voice Copilot is not.
Raw terminal TTS
It does not read every line aloud. It summarizes the important events.
Another agent
The Supervisor never writes code. It watches the one that does, and hands the decision to you.
Capabilities
Built for real coding sessions.
Supervisor+ pauses the agent
On a hard stop the wrapped CLI is suspended, you hear why, and the panel shows a Resume button. A watchdog makes sure a silent browser never wedges it.
Automatic model tiers
Weakest model for the narrator, strongest for the Supervisor, read from the CLI's own catalog. Explicit picks always win.
Hold while a line is read
Optional: the agent works in bursts between sentences instead of running ahead while you listen.
Structured CLI adapters
Stream-json where CLIs expose events, an HTTP proxy where they do not - including Codex on a ChatGPT plan - and PTY fallback when needed.
Pluggable providers
Reuse the CLI you launched with no keys, or choose cloud or local TTS, STT, and LLM backends. Provider failures surface loudly.
Trace when you need it
Listening is the default, but the popup keeps a readable trace with every verdict, pause, and prompt.
Alpha install
Try it with your CLI agent.
Voice Copilot is an early alpha for advanced users who are comfortable testing CLI workflows and sharing practical feedback. Turn the Supervisor on under Settings, Narration detail.
pipx install voice-copilot
# wrap the CLI you already use - narrator and
# supervisor reuse its login, no keys needed
vc codex
vc claude
# or with uv
uvx voice-copilot codex