Your meeting notes are only useful if you can find the right detail at the right moment. The Model Context Protocol (MCP) turns that search into a conversation: instead of scrolling through transcripts, you ask your AI assistant — "what did we promise the client last Tuesday?" — and it reads the answer straight from your recordings.
This guide shows how to connect Speak-Y's built-in MCP server to Claude, Cursor, ChatGPT and other MCP clients, exactly what the assistant can read and what it can change, and prompts that work well in practice.
The Model Context Protocol is an open standard that lets AI applications talk to external data sources through a common interface. An MCP server exposes data — in this case your meeting recordings — and an MCP client (Claude, Cursor, ChatGPT Desktop and dozens of others) consumes it. One server, every assistant: you configure access once and every MCP-compatible tool on your machine can use it.
Speak-Y ships with an MCP server built into the desktop app. It offers thirteen tools, in two groups that behave differently.
Five tools read, and they work on their own. The server opens the recording library on your Mac directly, so these answer even when the Speak-Y app is not running:
Eight tools act, and they run inside the app. Changing a recording, or touching an end-to-end encrypted team channel, needs the app itself: it holds the database and the encryption keys. With Speak-Y running, an assistant can tag a recording, put a person's name on a speaker in a meeting, rename a recording or mark it a favourite, transcribe a dictation again, list your team channels, create a channel in an existing workspace, publish a recording to a channel, and pull in recordings made on your phone or another Mac before answering.
Nothing in that list deletes anything: there is no delete tool at all.
Three things make this setup different from cloud notetakers:
For clients without one-click install, the same screen shows the manual configuration snippet to paste into the client's MCP config file. The MCP documentation covers every supported client, manual configuration and troubleshooting.
Once connected, the assistant decides when to call the Speak-Y tools. These patterns work reliably:
The common thread: name what you are looking for (topic, person, timeframe) and say what you want done with it. The assistant chains search → read → act on its own.
Yes, and the line is worth stating precisely, because "read-only" is the easy answer and it is no longer the true one.
Three mechanisms sit between the model and your library, and they are structural rather than a promise:
The privacy question matters more here than with most integrations, because meeting recordings are some of the most sensitive data on your machine. Speak-Y's answer is structural rather than contractual: reading is local, transcripts stay on your device by default, and audio is deleted after processing. The assistant reads exactly what you could read in the app yourself — no more — and it leaves the machine only for what you asked for: syncing your own devices, or publishing a recording to a team channel, both end-to-end encrypted. Details are in the privacy policy.
One limit no local server removes: whatever the assistant reads becomes part of your conversation and travels to your model provider like the rest of the chat. That distinction is unpacked in what an MCP server actually is.
Connecting your own notes is step one. If your team shares meetings into a Speak-Y team workspace, the same conversational access applies to the knowledge your team has decided to share — with end-to-end encryption handling who can see what. That combination — searchable team memory plus an AI assistant that can read it, and file into it when you say so — is what turns meeting notes from an archive into something you actually use every day.
Reading is local. The Speak-Y MCP server opens the recording library on your own machine, and nothing is uploaded so that an assistant can read it. Data leaves the Mac only for actions you ask for — syncing your other devices, or publishing a recording to a team channel — and both are end-to-end encrypted.
No. The built-in MCP server is free on every Speak-Y plan, including the Free plan. Most meeting notetakers that offer MCP gate it behind paid cloud tiers.
Any MCP-compatible client: Claude Code, Claude Desktop, ChatGPT Desktop, Cursor, Windsurf, VS Code, Zed, JetBrains AI Assistant, Raycast, Warp, Cline, Continue, Gemini CLI, LM Studio and others.
It can edit some things and delete nothing. Five tools read; eight more can tag a recording, name a speaker, rename it, transcribe a dictation again, create a team channel or publish a recording into one. Those are declared to the client as data-changing, so your assistant asks before calling them, and every action is logged in the app. No tool deletes a recording.
Yes. Settings → Integrations has a switch for assistant actions. Turned off, the eight action tools refuse and the five read tools keep working, so search and transcripts stay available with nothing writable.