MCP SERVER & CLI
Your conversations, inside your AI assistant.
The official VYSLO MCP server lets Claude, ChatGPT, Cursor and other MCP clients read your recordings (transcripts, summaries, speaker identity and the original audio) once you approve them, scope by scope. Ask about last week's conversations in the tool you already work in, with no export step and no copy-paste.
VYSLO 官方 MCP server 讓 Claude、ChatGPT、Cursor 等 MCP 客戶端，在你逐項授權之後讀取你的錄音——轉寫、摘要、 說話人身份，以及原始音頻。在你本來就在用的工具裡直接問上週的對話，不用導出，也不用複製貼上。
What you can ask
"Who did I promise what to last week?"
Answered from to-dos and speaker identity, listing only the commitments you made yourself, each with its recording and moment.
"Find the time Alex and I talked about the budget."
A search across all your transcripts, returning the passage, the recording and the time offset.
"How much of this week's talking was me?"
Weekly insights: your speaking share, talk time and recurring topics, compared with last week.
「上週我答應了誰什麼？」
由待辦與說話人身份回答，只列出你自己做出的承諾，每條都附來源錄音與時間點。
「找出我上次和 Alex 聊預算的那段。」
跨全部轉寫搜索，返回命中片段、所在錄音與時間偏移。
「這週我說話佔了多少？」
週洞察：你的發言佔比、說話時長與高頻主題，並與上週對比。
Not only what was said, but what you said
Transcripts and summaries are the baseline. Because the card knows whose voice it is, the VYSLO MCP server opens four more layers on top of them:
轉寫與摘要是基礎。因為卡片知道是誰的聲音，VYSLO MCP server 在此之上再開放四層：
Speaker identity
Each speaker marked as you, likely you or not you, together with the source of that verdict. Voiceprint templates never leave the platform.
Original audio
A short-lived signed link to the recording itself, behind its own consent, for clients that want to listen or run their own processing.
Weekly insights
Speaking share, talk time, recurring topics and to-dos, centred on you rather than on the meeting.
Card status
Battery, storage, firmware and last sync, returned with the time they were observed. A snapshot never poses as live.
說話人身份
每位說話人標為「是你」「可能是你」「不是你」，並附這個判決的來源。聲紋模板永不離開平台。
原始音頻
錄音本身的短期簽名鏈接，需單獨授權。適合想直接聽、或要自己做處理的客戶端。
週洞察
發言佔比、說話時長、高頻主題與待辦——以你為中心，而不是以會議為中心。
卡片狀態
電量、存儲、固件版本與最近同步時間，附帶觀測時刻。快照不會冒充實時狀態。
Connect
Works with Claude Desktop, Claude Code, ChatGPT, Cursor, Windsurf, Gemini CLI and Zed, and with any other client that speaks MCP. There are two ways in, and both call the same public API:
支援 Claude Desktop、Claude Code、ChatGPT、Cursor、Windsurf、Gemini CLI 與 Zed，以及其他任何支援 MCP 的客戶端。接入方式有兩種，調用的都是同一套公開 API：
| Mode | How it connects | Sign-in |
|---|---|---|
| Remote | Your client connects to the VYSLO server URL over Streamable HTTP | OAuth in the browser, approved in the VYSLO app or on the web |
| Local | Your client starts vyslo mcp as a local process over stdio | vyslo login once; the CLI holds the token |
| 方式 | 如何連接 | 登入 |
|---|---|---|
| 遠程 | 客戶端經 Streamable HTTP 連接 VYSLO server 地址 | 瀏覽器 OAuth，在 VYSLO App 或網頁上同意 |
| 本地 | 客戶端以 stdio 啟動本地進程 vyslo mcp | 執行一次 vyslo login，令牌由 CLI 保存 |
Remote: add the server by URL
In Claude Desktop, Claude on the web and ChatGPT, add a custom connector with this URL. The first tool call opens the consent page.
在 Claude Desktop、Claude 網頁版與 ChatGPT 裡，用這個地址添加自定義連接器。第一次調用工具時會打開授權頁。
https://api.vyslo.ai:8443/mcpclaude mcp add --transport http vyslo https://api.vyslo.ai:8443/mcp{
"mcpServers": {
"vyslo": { "url": "https://api.vyslo.ai:8443/mcp" }
}
}Local: through the vyslo CLI
Sign in once with vyslo login, then point any client that starts local servers at
vyslo mcp.
先執行一次 vyslo login 登入，再讓任何以本地進程啟動 server 的客戶端指向 vyslo mcp。
vyslo login
claude mcp add vyslo -- vyslo mcp{
"mcpServers": {
"vyslo": { "command": "vyslo", "args": ["mcp"] }
}
}Tools
| Tool | Returns | Scope |
|---|---|---|
list_recordings | Recordings in a time range, filterable by keyword or by whether you speak in them | recordings.read |
get_recording | Metadata and transcript segments, each with its speaker identity | recordings.read |
get_summary | Title, key points, to-dos and decisions | recordings.read |
search_transcripts | Matching passages across recordings, with recording and time offset | recordings.read |
list_speakers | Each speaker as you, likely you or not you, with the source of the verdict | recordings.read |
export_recording | The recording as Markdown, plain text, JSON or SRT | recordings.read |
get_audio_url | A short-lived signed link to the original audio, optionally for a time range | recordings.audio |
get_weekly_insights | The week centred on you: talk time, speaking share, recurring topics, compared with last week | insights.read |
get_speaking_stats | Speaking time and share by day or week | insights.read |
get_todos | To-dos with their source recording, time offset and whether you made the commitment | insights.read |
list_devices | Cards on the account: name, model, firmware, last seen | device.status.read |
get_device_status | Battery, storage, charging, smart wake, last sync, with observedAt | device.status.read |
| 工具 | 返回 | 權限 |
|---|---|---|
list_recordings | 時間範圍內的錄音，可按關鍵詞或「是否有你發言」過濾 | recordings.read |
get_recording | 元數據與轉寫分段，每段附說話人身份 | recordings.read |
get_summary | 標題、要點、待辦與決議 | recordings.read |
search_transcripts | 跨錄音的命中片段，附所在錄音與時間偏移 | recordings.read |
list_speakers | 每位說話人：是你／可能是你／不是你，附判決來源 | recordings.read |
export_recording | 以 Markdown、純文本、JSON 或 SRT 導出錄音 | recordings.read |
get_audio_url | 原始音頻的短期簽名鏈接，可指定時間區間 | recordings.audio |
get_weekly_insights | 以你為中心的一週：說話時長、發言佔比、高頻主題，並與上週對比 | insights.read |
get_speaking_stats | 按日或按週的發言時長與佔比 | insights.read |
get_todos | 待辦列表，附來源錄音、時間偏移與是否為你本人的承諾 | insights.read |
list_devices | 帳號下的卡片：名稱、型號、固件版本、最近在線 | device.status.read |
get_device_status | 電量、存儲、充電、智能喚醒、最近同步，附 observedAt | device.status.read |
Every tool reads. None of them starts a transcription or a summary, and nothing destructive (deleting recordings, unbinding a card, resetting a voiceprint) is exposed through MCP. Each scope is a separate tick on the consent page; see Scopes & tokens.
所有工具都只讀：不會發起轉寫或摘要；破壞性操作（刪除錄音、解綁卡片、重置聲紋）不經 MCP 開放。 每個權限在授權頁上單獨勾選，見權限與令牌。
How authorization works
- Add the VYSLO server in your client. A browser window opens the VYSLO consent page; it hands off to the VYSLO app if it is installed, otherwise you sign in on the web.
- The consent page names the client and what it asks for, for example "Claude Desktop wants to read your recordings and transcripts, and your weekly insights". You tick each item.
- The client receives the same capability token as any other app: 60 days, renewed silently.
- Each MCP client shows up on its own in the VYSLO app under 我的 → 已授權的應用 and can be revoked by itself. The server refuses the next call after a revocation.
- 在客戶端裡添加 VYSLO server。瀏覽器會打開 VYSLO 授權頁；裝了 VYSLO App 就跳轉 App，沒裝就在網頁登入。
- 授權頁寫明是哪個客戶端、申請什麼，例如「Claude Desktop 申請：讀取錄音與轉寫、讀取週洞察」。你逐項勾選。
- 客戶端拿到的是和其他應用一樣的能力令牌：60 天有效，靜默續簽。
- 每個 MCP 客戶端都會在 VYSLO App「我的 → 已授權的應用」裡單獨列出，可以單獨撤銷。撤銷之後，服務端立刻拒絕下一次調用。
The MCP server is an ordinary client of the public platform, with no private channel. Everything under Four commitments applies to it unchanged.
MCP server 是開放平台的普通客戶端，沒有私有通道。四條承諾對它同樣成立。
Privacy
- Only recordings marked
visibility=visibleare ever returned. Hidden or rejected recordings never appear, under any tool or any filter. - Identity results carry a verdict and its source. Voiceprint templates, embeddings and raw scores are never exposed.
- Audio links are signed, expire within minutes, and are issued only under
recordings.audio, which the user approves separately from reading text. - Transcript text is returned as data, separated from structured fields and labelled as the user's recording. The server never acts on instructions that appear inside a transcript.
- Long transcripts are paged and output is capped per call, so no tool hands over your whole history at once.
- Every tool call is logged per client, so the VYSLO app can show which AI tool read what, and when.
- 只有
visibility=visible的錄音會被返回。隱藏或被拒的錄音，在任何工具、任何過濾條件下都不會出現。 - 身份結果只給判決與來源。聲紋模板、embedding 與原始分數永不暴露。
- 音頻鏈接帶簽名、數分鐘內過期，且只在
recordings.audio之下簽發——這個權限與讀取文字分開，由用戶單獨同意。 - 轉寫正文作為數據返回，與結構化欄位分開，並標明是用戶錄音內容。server 不會執行轉寫裡出現的任何指令。
- 長轉寫分頁返回、單次輸出有上限，沒有任何工具會一次倒出你的全部歷史。
- 每次工具調用都按客戶端記錄，VYSLO App 裡可以看到哪個 AI 工具在什麼時候讀了什麼。
vyslo CLI
A single binary for macOS, Windows and Linux, on the same public API as the SDK and the MCP server.
macOS、Windows、Linux 單一可執行檔，與 SDK、MCP server 使用同一套公開 API。
| Command | Does |
|---|---|
vyslo login | Sign in through the browser |
vyslo recordings list | get | search | export | List, read, search and export recordings |
vyslo speakers <id> | Speaker identity for one recording |
vyslo insights weekly | This week's insights |
vyslo devices list | status | Cards on the account and their status |
vyslo mcp | Run the MCP server over stdio for local clients |
vyslo webhook listen | Receive sandbox webhooks locally while you develop |
| 命令 | 作用 |
|---|---|
vyslo login | 經瀏覽器登入 |
vyslo recordings list | get | search | export | 列出、讀取、搜索、導出錄音 |
vyslo speakers <id> | 單條錄音的說話人身份 |
vyslo insights weekly | 本週洞察 |
vyslo devices list | status | 帳號下的卡片及其狀態 |
vyslo mcp | 以 stdio 為本地客戶端運行 MCP server |
vyslo webhook listen | 開發時在本機接收 sandbox webhook |
Questions
A client that will not connect, a tool that returns something unexpected, or a question you wish your assistant could answer: write to support@vyslo.ai and say which client you use.
客戶端連不上、工具返回不符合預期，或你希望助手能回答某類問題：寫信到 support@vyslo.ai，並註明你用的是哪個客戶端。