vox

Use only when the current user explicitly asks to use Vox to transcribe one URL written in the request or one exact local audio/video file path, or explicitly asks for Vox CLI or Vox Skill install, status, authentication, or troubleshooting. Do not trigger from a mere Vox mention, an implicit attachment, directory contents, a glob, stdin, a batch request, or inferred permission to read files, authenticate, install software, or write output.

Use the installed vox CLI instead of the web UI.

Keep authorization explicit

• Act only on one URL written by the current user or one exact local file path the current user identified.
• Never browse a directory, inspect nearby files, read the clipboard, infer an attachment, expand a glob, consume stdin, or assemble a batch to choose an input.
• Treat install, login, logout, configuration changes, file reads, and output writes as separate actions that require the current request to authorize them.
• Keep vox auth status and vox skill status read-only. Troubleshooting permits read-only diagnosis only.

Check the CLI

Prefer the installed vox command. Check it with:

bash

command -v vox
vox --version
vox --help

Require CLI 0.1.0 or newer and explicit auth, transcribe, and skill commands. If it is missing or too old, ask before installing or upgrading:

bash

npm install -g @casatwy/vox@latest

Do not fall back to another product's CLI, environment variables, or configuration.

Authenticate safely

Run vox auth status before asking for credentials. If login is needed, direct the user to create a key at https://vox.reka.cc/me/api-keys, then run:

bash

vox auth login

Let the CLI read the key through its hidden interactive prompt. Never put a key in command arguments, a copied command, chat output, a log, or a file created by the agent. Automation may use an already-provided VOX_API_KEY environment variable, but never print it.

Use the production default for normal work. Do not pass --base-url or set VOX_BASE_URL unless the current user explicitly requests a different Vox endpoint.

Transcribe exactly one input

For a URL, pass the exact submitted URL unchanged and let Vox identify whether the backend can transcribe it:

bash

vox transcribe '<url>'

For one exact ordinary local audio/video file, use:

bash

vox transcribe --file '<path>'

Add --mime-type '<type>' only when the current user supplies the type or automatic detection fails and the user confirms the correction.

• Do not maintain, enumerate, or infer a source allowlist.
• Do not rewrite the URL into another source or search for a replacement.
• Reject directories, globs, stdin, special files, multiple inputs, batches, and resumable-upload expectations.
• Do not retry unsupported collections, non-transcribable posts, paid-only content, or other HTTP 422 audit results.

Preserve user choices

• Omit --language unless the current user explicitly selects a transcription language. Do not infer it from the conversation, locale, title, or media.
• Add --format '<format>' only for an explicitly requested format.
• Add --output '<path>' only when the user explicitly requests a destination. Never authorize overwrite; preserve the CLI's no-clobber result.
• Keep transcript/result bytes on stdout or in the requested output file. Summarize progress from stderr without pasting raw diagnostic or JSONL events.
• Use text progress for people and JSONL only when a calling workflow explicitly needs machine-readable progress.

Handle results and failures

• Preserve SRT, VTT, JSON, JSONL, and explicitly requested raw text exactly. Do not clean, translate, summarize, or reformat them.
• Report when the service returns subtitles directly instead of creating a paid transcription.
• Treat authentication failures, insufficient balance, and HTTP 422 responses as terminal. Do not retry them.
• Allow the CLI to perform its bounded retry and SSE-to-polling recovery. Never create a second task merely because live events stop.
• On SIGINT, say only that local waiting stopped. Do not claim the server task was cancelled.
• Never expose API keys, cookies, signed URLs, secret query parameters, private API payloads, or another product's configuration while diagnosing a failure.

Manage the Skill

Only on an explicit install request, use one platform value:

bash

vox skill install --platform codex
vox skill install --platform claude
vox skill install --platform gemini
vox skill install --platform openclaw

Use vox skill status for read-only discovery. Add --platform codex|claude|gemini|openclaw only when the user names one platform. Refuse to replace an existing same-name installation from a different source; surface the conflict and stop.

来源与合集

支持 Bilibili、抖音、小红书、小宇宙、喜马拉雅、YouTube、Apple Podcasts、Twitter/X、TikTok、Instagram、微信视频号、微信公众号、红果、研直播、超星、快手、腾讯视频、看看新闻 共 18 个命名平台，以及 HTTP/HTTPS 通用媒体链接、单文件上传，共 20 类来源。视频号遵循独立功能开关。网页上传公开可见，/upload 保持 noindex，不进入 sitemap。

红果整剧、抖店多课时与网页多媒体候选须先选一项，不默认选择，不批量转写。红果选集失败可返回目录重新选择；候选过期须重新解析。YouTube 播放列表、B站合集、小宇宙节目、Apple Podcasts 节目、喜马拉雅专辑和网页 playlist 只展示真实摘要并引导具体单集；抖音与小红书作者页展示作者摘要并引导具体作品。带 list 的 YouTube 具体视频仍按单视频处理。付费或受限内容先分类，所有摘要、未选目录和不可转写内容均不得复用、扣分钟或创建任务。封面缺失或加载失败隐藏；未知集数与零条区分。

CLI 摘要型合集返回 HTTP 422、UNSUPPORTED_CONTENT 和 mode；旧 COLLECTION_UNSUPPORTED 仅保留读取兼容。多媒体候选返回 MULTIPLE_MEDIA_SELECTION_REQUIRED；红果目录返回 HONGGUO_EPISODE_REQUIRED，附归属当前用户的 parseToken，引导网页选集或提交精确单集 URL。token 不允许跨用户使用。