Natively MCP server
Natively includes a local MCP server that lets AI apps on your computer, such as Claude Desktop, Claude Code, Cursor or Codex, list, search and read the meetings saved in Natively. It is off by default, listens only on 127.0.0.1, needs a token that Natively generates, and is read-only: its four tools can read notes and transcripts but cannot change or delete anything.
By Evin JohnPublished 6 min read
What MCP is and why Natively has a server for it
The Model Context Protocol (MCP) is a common way for AI apps to call tools that live outside them. A coding assistant or chat app that speaks MCP can ask a server what tools it offers, then call them in the middle of a conversation. Natively's server offers one thing: your meeting history. With it on, you can ask Claude what was decided in last Tuesday's design review, or have Cursor read the transcript of a planning call while it writes the code that came out of it.
Natively keeps meetings on your computer rather than on a server, so the MCP server runs inside the desktop app too, instead of at a public web address. That shapes everything below: it works only on the machine where Natively is installed, only while Natively is open, and only for apps on that machine.
The four tools
| Tool | What it returns | Inputs |
|---|---|---|
list_meetings | Meetings newest first, with id, title, date and length | Optional from and to dates (ISO 8601), optional query matched against title and notes, limit 1 to 100 (default 20) |
get_meeting_notes | One meeting's notes as Markdown: summary, sections and action items, with the date, the attendees and your own typed notes | id from a list or search |
get_transcript | One meeting's transcript, one line per utterance with the speaker and the time from the start | id, optional from_seconds and to_seconds for part of the meeting |
search_meetings | Meetings whose title, notes or transcript mention the words, ranked, with short excerpts | query, limit 1 to 50 (default 10) |
Lists and searches look through your 1,000 most recent meetings. A very long transcript is cut at about 200,000 characters per call, and the server tells the AI app to ask for a later time range for the rest. Attendees include their email addresses when the meeting was linked to a calendar event. A bad input, such as an unknown meeting id, comes back as a plain sentence the model can read and correct.
Turn it on
- Open the MCP cardIn Natively, go to Settings > Connectors and scroll to MCP. While it is off, the card says: Off. Nothing can read your meetings.
- Press Turn onNatively starts the server on 127.0.0.1 at a port it picks once and keeps, so the address you paste stays valid. The status changes to: On. Only apps on this computer can connect.
- Pick your appThe card shows a tab for Claude Code, Claude Desktop, Cursor and Codex, each with its setup already filled in with your address and token. Copy it.
- Paste it where the card saysClaude Code: run the command in a terminal. Claude Desktop: add it to claude_desktop_config.json. Cursor: add it to mcp.json. Codex: add it to config.toml in your .codex folder. Restart the app if it was already open.
- Ask about a meetingIn the AI app, ask a question that needs your meetings. The app decides when to call Natively's tools.
What the setup looks like
These are the shapes Natively generates. In the app, <port> and <token> are already your real values; you never need to type them yourself.
claude mcp add --transport http natively http://127.0.0.1:<port>/mcp --header "Authorization: Bearer <token>"{
"mcpServers": {
"natively": {
"url": "http://127.0.0.1:<port>/mcp",
"headers": { "Authorization": "Bearer <token>" }
}
}
}[mcp_servers.natively]
url = "http://127.0.0.1:<port>/mcp"
http_headers = { Authorization = "Bearer <token>" }Claude Desktop is set up differently, so copy its tab rather than adapting the Cursor block. It reaches HTTP servers through the mcp-remote bridge, which its config starts with npx -y mcp-remote followed by the address and --allow-http, so Node.js has to be installed for that one. The token travels in an environment variable named NATIVELY_AUTH on purpose: Claude Desktop on Windows does not handle spaces inside arguments, and this form works on both macOS and Windows. The other three apps talk to the server directly.
Security: who can reach it
- Off until you turn it on. A fresh install has no server listening. Turning it off stops the server. If you leave it on, it starts again the next time Natively opens.
- This computer only. It binds to 127.0.0.1, the loopback address, and the address is not configurable. Other devices on your network cannot connect.
- A token on every request. Each request needs the bearer token from Settings, compared in constant time. The token is stored encrypted on your computer, with the macOS Keychain on a Mac and Windows' data protection (DPAPI) on Windows.
- No web pages. A request with a browser origin or a host name that is not loopback is refused, which blocks a website from reaching the server through DNS rebinding.
- Read-only by design. There is no tool that writes, edits or deletes. Request bodies are capped at 1 MB.
- Only while Natively is open. The server runs inside the app and stops when you quit, on both macOS and Windows.
What it can and cannot see
| Can read | Cannot read or do |
|---|---|
| Meeting titles, dates and lengths | Audio, screenshots or attached files |
| Notes: summary, sections, action items | Your settings, API keys or connector tokens |
| Your own typed notes on a meeting | Anything outside Natively's meeting history |
| Transcripts with speaker names and times | Create, edit, move or delete a meeting |
| Invitee names and emails from the calendar event | Send anything to Slack, a CRM or email |
The server itself never sends anything off your computer. The AI app that calls it is a different matter: Claude Desktop, Claude Code, Cursor and Codex normally send the conversation, including whatever a tool returned, to their own cloud model. Reading a transcript into one of them shares it with that provider under that provider's terms. If your meetings must stay on the machine, keep the server off or use it only with an AI app that runs a local model.
Fictional example
- In Claude Desktop
- "Search my Natively meetings for 'pricing page' from October and list every decision we made, with the meeting it came from."
- In Cursor
- "Read the transcript of yesterday's API design review in Natively and check whether this migration matches what we agreed about pagination."
- In Claude Code
- "Get the action items from my last three one-on-ones with Priya and draft an agenda for Thursday."
- In Codex
- "Find the meeting where we discussed the retry policy and quote what Omar said about timeouts."
Precise prompts help: name a person, a topic or a date range, so the app calls search_meetings or list_meetings with a narrow query before reading whole transcripts.
Plans and platforms
The MCP server is part of the free app, on macOS and on Windows, and behaves the same on both. It does not run on Linux, because Natively does not. To send notes to other tools rather than letting AI apps read them, see the integrations page; for keeping the whole pipeline on your machine, see private meetings with local AI.
Questions people ask
Is the Natively MCP server on by default?
No. It stays off until you press Turn on in Settings > Connectors, under MCP. It then listens only on 127.0.0.1 and needs the token Natively generates.
Can an AI app change my meetings through MCP?
No. The four tools (list_meetings, get_meeting_notes, get_transcript, search_meetings) only read. Nothing can be created, edited or deleted through the server.
Does my meeting data leave my computer when I use it?
The server sends nothing anywhere, but the AI app that calls it usually passes what it read to its own cloud model. Keep the server off, or use an app with a local model, if transcripts must stay on the machine.
Where to go next
Natively runs on macOS and Windows. The download is free; Pro and the Natively API are optional.
Facts on this page were last checked on . Product details come from the Natively app itself; if something here is out of date, tell us at hello@heynatively.com.
