Security and permissions
How does an assistant get access?
Section titled “How does an assistant get access?”Meetingize implements the MCP authorization specification:
- The assistant calls the server without a token and receives
401with a link to the protected resource metadata (RFC 9728). - It discovers the Meetingize authorization server (RFC 8414), registers itself (dynamic client registration, RFC 7591, or a Client ID Metadata Document) and opens the Meetingize consent screen in your browser, with PKCE (S256).
- You sign in to Meetingize (the assistant never sees your credentials), check the application name, its domain and the permissions requested, and click Allow access or Deny.
- The assistant receives an access token valid 1 hour and a refresh token valid 30 days, rotated at each use. Tokens are bound to your account and to the scopes you approved.
Permissions (scopes)
Section titled “Permissions (scopes)”| Permission | On the consent screen |
|---|---|
read:meetings | Read your meetings (list, search, details, summary, participants, templates) |
read:transcripts | Read the transcripts of your meetings |
read:documents | Read the documents generated from your meetings |
read:usage | Read your API consumption and wallet balance |
write:documents | Generate new documents from your meetings (uses your API allowance) |
write:meetings | Create meetings from audio URLs (uses your API allowance) and update their title, date, tags and participants |
Write permissions are highlighted on the consent screen. The MCP server can never delete a meeting or a document,
nor change your account or your sharing settings: no tool does it. With write:meetings it can only rename a
meeting, change its date and tags, and set who the speakers are, after your confirmation.
Isolation
Section titled “Isolation”Every MCP call is executed by the same authorization layer as the public API. A token only reaches the data of the user who approved it; meetings of other accounts answer “not found”. Meetings shared with you by someone else are not exposed to assistants.
Revoking access
Section titled “Revoking access”Settings > Connected applications in the web app lists every assistant you connected, its permissions, when you authorized it and when it was last used. Disconnect deletes its tokens immediately; the assistant must ask for your consent again to reconnect.
Write confirmations and costs
Section titled “Write confirmations and costs”Tools that consume your API allowance or change data (generate_document, create_meeting, update_meeting,
send_feedback) only act on a second call with
confirm: true, after showing you a preview. Your assistant (Claude, ChatGPT) also asks for your approval before
running non read-only tools. Usage alerts and a spend cap protect you from surprises, see
billing.
Good practices
Section titled “Good practices”- Approve only the permissions you need. A read-only connection (
read:*) is enough to ask questions about your meetings. - Treat answers as drafts: quote-check important facts against the transcript (the tools return timestamps for that).
- If you use an API key in a configuration file (Cursor, Claude Code), keep it out of version control and give it read-only scopes.
- Remember that your assistant’s provider processes what the assistant reads (transcripts, reports) under its own terms. Do not connect accounts holding data you are not allowed to share with that provider.