VOICEDROP CONNECTIONS
Bring your history into your AI
One read-only MCP server: https://tryvoicedrop.com/mcp. Each AI client has its own connection and consent. Cloud history and optionally synced dictations are available; raw audio and unsynced local dictations are not.
ChatGPT: create a custom plugin
- Use ChatGPT on the web. Open Plugins → + → Add custom MCP server. Older interfaces may use Settings → Apps → Advanced settings / Developer mode → Create.
- Name: VoiceDrop. Server URL:
https://tryvoicedrop.com/mcp. Choose OAuth and dynamic client registration (DCR) when asked. Leave static client credentials empty. - Create/install the custom plugin, sign in to your VoiceDrop account and approve read-only access. Check the account shown before approving.
- Enable or @mention VoiceDrop in a new chat. Verify its tools are available, then ask it to find a meeting only when you want to share that history.
Custom MCP availability depends on your account and workspace permissions. This is a personal/workspace custom plugin, not a public directory listing. No API key is needed. Access lasts 30 days; reconnect after expiry.
OpenAI’s current custom MCP guide
Codex on Windows or Mac
Codex uses remote HTTP MCP and browser OAuth. Inspect only the VoiceDrop entry and preserve other servers. Run the following in an interactive PowerShell or Terminal window:
codex mcp get voicedrop # Only if missing: codex mcp add voicedrop --url https://tryvoicedrop.com/mcp # Only if authentication is needed: codex mcp login voicedrop --scopes history:read
Keep the login window open and approve access in your browser. Use one authentication attempt at a time; never paste a token or callback URL into chat. Restart Codex if the tools are missing. Confirm voicedrop_list_history and voicedrop_get_history before reporting a connection as successful, and ask before reading private history.
Other compatible MCP clients
Use the same URL with a client that supports Streamable HTTP MCP, browser OAuth and dynamic client registration. Local clients use a localhost callback; hosted callbacks currently support ChatGPT and Claude Chat. Grok Build documents remote HTTP MCP, but its VoiceDrop login has not been verified. Other products, including Muse, need their own compatibility check. A model’s name alone does not establish connector support. The optional stdio bridge below provides another route for clients with local MCP support.
Claude Chat on Windows or Mac
Use claude.ai in your browser or the Claude desktop app. You do not need a terminal, API key or local bridge.
- Open Claude → Customize → Connectors. Some versions use Settings → Connectors.
- Choose + → Add custom connector. Name it VoiceDrop and enter
https://tryvoicedrop.com/mcp. Leave advanced OAuth fields empty. - Choose Connect, sign in to VoiceDrop and approve read-only access. Check that the account shown is yours.
- Start a new chat and enable VoiceDrop under + → Connectors. Ask Claude to find your latest meeting and summarize the decisions.
Claude can read cloud transcripts, speaker information and Smart Notes. Quick dictations are included only after you explicitly enable cloud sync. Raw audio is not exposed. Access expires after 30 days; reconnect then. Revoke access sooner in VoiceDrop → Settings → AI connections. Organization owners may need to enable custom connectors.
Claude Code: let your agent set it up
Claude Code has its own MCP configuration. Your Claude Chat connection does not automatically configure it. Paste this message into Claude Code:
Connect VoiceDrop in Claude Code using remote HTTP MCP and browser OAuth. Inspect only the VoiceDrop entry with: claude mcp get voicedrop. If absent, run: claude mcp add --transport http --scope user voicedrop https://tryvoicedrop.com/mcp Keep an existing HTTP entry for this URL. Repair only a stale VoiceDrop entry; preserve other MCP servers. If already connected, skip authentication. Otherwise guide me to type /mcp in this Claude Code session, select VoiceDrop and choose Authenticate. If the newly added server is missing, reopen the session first. I will sign in and approve read-only access in the browser. Do not attempt login through a non-interactive shell or a ! command. Alternatively, give me claude mcp login voicedrop to run in a separate interactive PowerShell/Terminal window, kept open until completion. Use one login attempt at a time; do not replace pending login or reuse an old consent link. Never ask for passwords, keys, tokens or callback URLs in chat. For a localhost error, use the original login's own redirect prompt, or cancel before starting fresh authentication. Verify with claude mcp get voicedrop. If tools are missing, restart the current session or reconnect through /mcp. Verify voicedrop_list_history and voicedrop_get_history are available, then ask me before reading private history.
- Let Claude add VoiceDrop using the message above.
- Type
/mcpin Claude Code → VoiceDrop → Authenticate. Sign in and approve access. Reopen the session first if VoiceDrop is missing. - Check that Claude reports Connected and can see both VoiceDrop tools. Already connected? Skip sign-in.
No API key, local bridge or client secret is needed. Alternatively, run these commands in a separate interactive PowerShell window on Windows or Terminal on Mac:
claude mcp add --transport http --scope user voicedrop https://tryvoicedrop.com/mcp claude mcp login voicedrop
If VoiceDrop is already added, skip the add command. Keep the terminal open until login finishes; avoid agent shell and ! commands. Use one authentication attempt at a time. If the browser shows a localhost error, use the redirect prompt in that same terminal, never paste its address into a chat. Cancel the old attempt before starting a new one. After approval, reopen your Claude Code session if the tools are missing. Read-only access lasts 30 days and appears as “Claude Code / local MCP” in VoiceDrop’s connection list.
Other AI clients: local MCP setup
This route is for clients that explicitly support local stdio MCP servers. Each client has its own configuration location and permission controls. It is separate from Claude Chat’s remote connector; Claude Chat and Claude Code use the browser OAuth setup above.
- Install Deno using the official Deno instructions, if it is not already installed.
- Download the VoiceDrop MCP bridge and save it in a permanent folder. On Windows, for example, use your Documents folder. On Mac, use a folder under your home directory.
- In VoiceDrop → Settings → AI connections → Other AI tools, create a read-only key. Enter it directly into the client’s private configuration or secret store. Never paste the key into a conversation.
- Adapt the configuration below to your client. Replace both absolute paths. On Windows, find Deno with
(Get-Command deno).Sourcein PowerShell; on Mac usecommand -v denoin Terminal. JSON Windows paths need doubled backslashes. - Restart or reload the MCP client and enable its VoiceDrop tools. The client should discover
voicedrop_list_historyandvoicedrop_get_history. Search both recording and quick_dictation collections when you want all available history.
{
"mcpServers": {
"voicedrop": {
"command": "ABSOLUTE_PATH_TO_DENO",
"args": [
"run",
"--no-config",
"--allow-env=VOICEDROP_HISTORY_URL,VOICEDROP_HISTORY_TOKEN",
"--allow-net=muocuuyvgyfczxuijgiu.supabase.co",
"ABSOLUTE_PATH_TO_VOICE_HISTORY_MCP_TS"
],
"env": {
"VOICEDROP_HISTORY_URL": "https://muocuuyvgyfczxuijgiu.supabase.co/functions/v1/history-api",
"VOICEDROP_HISTORY_TOKEN": "ENTER_KEY_PRIVATELY_IN_YOUR_CLIENT_CONFIG"
}
}
}
}The bridge uses Deno’s pinned npm dependencies and only permits network requests to VoiceDrop’s Supabase host. An AI coding assistant can prepare the paths and configuration for you, then let you enter the key privately. A chat interface without local MCP support cannot use this configuration.
Troubleshooting
In VoiceDrop, refresh the connection list to verify your Claude connection appears. If it cannot connect, check that your browser used the same VoiceDrop account. If Claude reports expired access, reconnect. If a dictation is missing, check the opt-in cloud sync setting. Revoking access prevents future reads but does not remove text already sent to your AI provider.