Security

Choose what meeting data your agent can read and where processing takes place. Zoom authorization, local storage and AI processing have separate boundaries: granting app access does not itself send a transcript to an AI service. This guide explains those boundaries before you connect or archive data.

CredentialsCopy Anchor Link

• zm uses a configured Zoom app with Public Client OAuth. Zoom can show a client secret for that app, but this tool never asks for or uses it. Configure only its Public Client ID.
• Login uses PKCE and checks the returned state. The callback page at wirecat.dev/zoom-callback is static and only passes the answer on to zm; the answer is useless without the verifier that stays on your machine.
• Tokens are stored only in the system keyring and are never written to a config file. A missing keyring is an error.
• The login listener closes on success, failure and timeout.

What reaches the networkCopy Anchor Link

• Zoom's API, for login, pull, recordings and remote meeting commands. --offline refuses network requests.
• A remote embedding endpoint only when you name it with --base-url (topic search).
• Metadata, help, configuration and recordings archive commands work without login.

What stays on diskCopy Anchor Link

• Meetings live in the shared local WireCat database (wirecat.db) unless MESSAGING_STORE points elsewhere.
• Recording manifests hold metadata, file sizes and SHA-256 checksums. Signed download URLs and OAuth tokens are excluded.
• Saved archive jobs capture the profile, public client ID, destination and paths, without credentials.
• Diagnostic messages exclude raw provider error text.

What agents can doCopy Anchor Link

• The MCP server is read-only, fixed to the selected local account, and exposes no file exports, remote mutations, embedding generation or task creation.
• Local audio transcription uses an already installed model on your machine and downloads none.
• Remote meeting changes, meeting links, search indexing, transcription, archive job creation, embedding generation and lock recovery preview by default and need --yes; --dry-run wins where offered. Remote writes are never automatically retried.
• Task proposals never create a task.

An AI app can send the tool results it reads to its own provider. Choose that app and its data policy separately from Zoom authorization. A remote embedding endpoint receives text for generation and queries only when explicitly selected; local model execution stays on your computer. Stored files and database content remain sensitive meeting data even though their manifests omit tokens.

To withdraw Zoom access, revoke the app authorization. To control what an AI app can request, choose the CLI skill or read-only MCP connection.