Cycling through all six. Tap any point to stop.

Measured on six things

Voice-driven supervision for AI coding agents—hears, warn...

Security

Probe found HTTPS not enforced, no HSTS, no CSP, no frame protection, and no Referrer-Policy — 2/12 checklist points, though SPF and DMARC are set. Reusing the CLI's own login avoids new key sprawl, but the site's own headers are thin.

Screenshots

1 / 1

Voice Copilot

About this app

Voice Copilot is a lightweight tool that wraps around your command-line AI coding agent (like Claude Code or Codex) to narrate its actions out loud and have a stronger model supervise its work. It reads the agent's event stream and converts it into short spoken updates using the cheapest model available, while a more capable model reviews the agent's progress at key checkpoints. When the agent drifts off task, the supervisor warns you verbally and can even pause the agent until you decide to resume or redirect. This lets you step away from the terminal without losing awareness, catching mistakes before they become costly. Built for developers who use LLM-powered coding assistants, Voice Copilot reuses your existing CLI login (no extra API keys) and supports multiple agents. The alpha release emphasizes real-time supervision with a focus on minimal token usage and a human-in-the-loop control.

Key features

1Cheap narrator model converts the CLI agent's event stream into short spoken updates in real time.
2Stronger supervisor model reviews at checkpoints and issues OK, WARN, or STOP verdicts.
3Supervisor+ pauses the wrapped agent on a hard stop until you resume from the panel or hotkey.
4Reuses your existing CLI login (Claude Code, Codex) with no extra API keys or second signup.
5Automatic model tiers: weakest for narration, strongest for supervision, read from the CLI's own catalog.

Target customer profile

Drawn from the product itself, not from a survey.

Demographic

Software developers using AI coding assistants

Pain points

Constantly monitoring the terminal; missing errors when multitasking

Primary needs

Away-from-keyboard awareness; real-time error prevention

Demographic

DevOps engineers managing automated scripts

Pain points

Unattended agents making destructive changes; lack of oversight

Primary needs

Automated supervision; pause-on-risk capability

Demographic

Tech leads overseeing multiple AI coding sessions

Pain points

Juggling multiple agents; ensuring compliance and quality

Primary needs

Scalable supervision; audible alerts for drift

Scout Review

Written by AI from measured evidence, scored out of 100.

Voice Copilot is a genuinely interesting idea wrapped in a competent but thin alpha. The two-model split — cheapest model narrates, strongest model supervises at checkpoints — is a smart answer to a real problem: you can't watch the terminal for ten minutes, and you shouldn't have to. The STOP-and-pause behavior with a Resume button is the kind of human-in-the-loop control that makes unattended agents actually usable. The landing page explains all of this clearly, with disciplined copy and a demo that respects your time. What holds it back is execution maturity: mobile paint times near ten seconds, a security header profile of 2/12, and an alpha install path aimed squarely at people who already live in a terminal. None of that is fatal for an early open-source tool, but it means the product is currently a promise with a good pitch rather than a proven daily driver.

Measure by measure

Usability

Zero-signup install that reuses your existing CLI login is genuinely low-friction, and the YouTube demo explains the loop fast. But it's an alpha for 'advanced users', and docs depth beyond the homepage couldn't be verified.

Design

Dark, terminal-native landing page with a clear 01-04 flow, a trace mockup, and a sharp 'what it is NOT' section. Copy is disciplined and on-brand for CLI devs; the actual popup UI is only shown as a static mockup.

Speed

Desktop Lighthouse lands 86/100, but mobile is 57/100 with a 9.3s LCP and 7.8s FCP — the page paints slowly on phones. TBT is a healthy 40ms and CLS is 0, so it's a paint problem, not a JS problem.

Security

Accessibility

Lighthouse accessibility scores 96/100 with a single failing audit: insufficient contrast between background and foreground colors. No a11y statement or multi-language support found, but the automated baseline is strong.

Growth

Real, specific pain — unattended agents drifting — with a concrete wedge: cheap narrator + strong supervisor at checkpoints and a STOP that pauses the agent. Multi-agent support and CLI-login reuse lower adoption friction.

Improvement areas

Security headers: HTTPS not enforced, no HSTS, CSP, or Referrer-Policy (2/12 probe points).
Mobile performance: 9.3s LCP and 7.8s FCP on Lighthouse mobile (57/100).
Contrast failure on dark theme; no accessibility statement or multi-language support.
Docs depth unverified beyond homepage; alpha install assumes advanced CLI users.

Design & Usability

Voice Copilot's site is built for its audience: dark, terminal-flavored, and structured as a 01-04 loop that explains narrator, supervisor, and human decision in four steps. The 'what it is NOT' section preempts the obvious misreads (it's not dictation, not raw TTS, not another agent), which is unusually disciplined copy. The weakness is that the product's real interface — the popup panel with trace, Resume, and hotkeys — only appears as a static mockup, so you're trusting the marketing to describe the UX. Usability is the stronger half: no signup, no new API keys, install and point it at your existing CLI, with a sub-minute YouTube demo. But it's explicitly an alpha for advanced users, and documentation depth beyond the homepage couldn't be verified from the collected evidence.

Speed & Security

Speed is a split verdict. Desktop Lighthouse lands at 86/100, which is fine for a marketing page, but mobile drops to 57/100 with a 9.3-second LCP and 7.8-second FCP — the page is light on JavaScript (TBT 40ms, CLS 0) yet slow to paint on phones, which is a hosting/asset problem rather than a code problem. For a product selling real-time awareness, a nine-second first paint is an awkward look. Security is the weakest measured dimension: the HTTP probe found HTTPS not enforced, no HSTS, no CSP, no frame protection, no X-Content-Type-Options, no Referrer-Policy, and no Permissions-Policy — 2 of 12 checklist points, with only SPF and DMARC passing. There's no security page, no SOC 2 or audit claim, and no bug bounty. The one genuine mitigant is architectural: it reuses your CLI's existing login instead of issuing new API keys, so credentials don't multiply. That's good design, but it doesn't fix the site's own header hygiene.

Accessibility & Growth

Accessibility is close to clean: Lighthouse scores 96/100 with exactly one failing audit — insufficient contrast between background and foreground colors, which tracks with a dark theme and muted grey secondary text. No accessibility statement, no multi-language support, and the CLI-plus-popup surface means keyboard support matters more than usual; the hotkey-driven Resume flow is at least a good sign. Growth is where the idea earns its keep. The problem it names is real and specific: agents run for minutes, and you either babysit the terminal or discover the bad turn after the fact. The wedge is concrete — a cheap model narrates, a strong model reviews only at checkpoints, and STOP actually pauses the agent — which is a token-budget argument incumbents like GitHub Copilot and Cursor don't make, because they're editors, not supervisors. Multi-agent support and CLI-login reuse lower adoption friction. The honest limit: it's an alpha 0.1.0 with no traction signals in the collected evidence, and its value depends entirely on how many developers run long unattended CLI sessions — a real but narrower segment than the editor-integrated giants serve.

Conclusion

If you run Claude Code or Codex for long stretches and have ever come back to find the agent three files deep in something you didn't ask for, this is aimed directly at you. Install it, watch the demo, and judge it on whether the Supervisor's WARN and STOP calls are actually right — that's the whole product. If you're a casual Copilot user who tabs in and out of an editor, this solves a problem you don't have yet. The idea deserves a real beta; the site and the security posture need to catch up to it.

Suggestions for improvement

Enforce HTTPS and add HSTS, CSP, and Referrer-Policy — 2/12 header checks is the easiest score to fix on this page.
Fix the mobile LCP: 9.3s first paint on a page selling real-time awareness undercuts the pitch.
Replace the static panel mockup with a real screenshot or short GIF of the trace, Resume button, and hotkey flow.
Publish a docs page covering install per CLI (Claude Code, Codex), model tier selection, and what STOP actually does.

How we score

Alternatives comparison

Named competitors, point by point. Nobody paid to appear here or to be left out.

	Voice Copilot	GitHub Copilot	Cursor
Voice narration of agent actions	Yes — cheap model narrates event stream aloud	No — silent inline suggestions	No — chat and inline edits, no narration
Second-model supervision with pause	Yes — Supervisor STOP pauses the agent	No — no supervisory pause	No — no multi-model oversight layer
Works with multiple CLI agents	Yes — Claude Code, Codex, other CLI agents	Editor/IDE extension, not a CLI wrapper	Single AI-first editor
Extra API keys required	No — reuses existing CLI login	Subscription-based, no separate key	Subscription-based, no separate key
Primary interface	CLI wrapper + browser popup panel with trace	IDE extension with inline suggestions	Full AI-first code editor