Let an AI agent read your stored meeting text and return source-backed answers. Choose a skill for a terminal agent or MCP for an AI app that launches local tools. A skill is an instruction file; MCP is the protocol the app uses to call tools. Neither connection logs in or downloads meetings. Import or pull at least one meeting before trying a transcript question.

A terminal agent can run the broader CLI, including writes when instructed. The MCP server provides a narrower read-only tool set. These are different access choices.

zm skill show prints the packaged instructions for an agent. zm skill install writes a version-stamped copy to the agent skills directories: ~/.claude/skills/zoom-cli/ for Claude Code and ~/.agents/skills/zoom-cli/ for Codex and Gemini CLI. --for claude or --for agents selects one; the default is both. Installation is explicit and does not run when a command starts.

zm serve --mcp uses MCP's stdio protocol and serves eight read-only tools: meeting list/show, transcripts, summaries, exports, word and participant searches, and evidence for project names. Tool discovery does not open a database. Each read uses the selected local account and closes its connection before the response; a caller cannot override the account. File exports and remote mutations are not exposed as tools.

MCP clients can launch zm with arguments ["serve", "--mcp"]. In the mcpServers format of Claude Desktop's configuration:

{
  "mcpServers": {
    "zoom": { "command": "zm", "args": ["serve", "--mcp"] }
  }
}

A desktop app may not see your terminal's PATH, and a stdio server inherits only part of the environment (MCP debugging guide). If the server does not start, use the full path from command -v zm on Linux/macOS or where.exe zm on Windows as the command. If you set ZM_CONFIG_DIR or MESSAGING_STORE, pass the same values in the entry's env, or the server reads a different profile or database.

Keep stdout for protocol messages; omit --json and --jsonl. --max-output-bytes bounds each tool result, with zero meaning unlimited. Select a profile and import or pull its data before reading tools; discovery is available without stored data.

meeting_people searches saved participant names/emails within the server-selected account. Use limit, scan_limit, max_read_bytes and cursor for bounded reads. The report's item.page and item.next step preserve continuation even when no observation matches the current scan page. identity_id identifies an observed identity; repeated names never establish person links. meetings_list takes a metadata query (for example started_at:[2026-09-12 TO 2026-10-12]) instead of its former since and until arguments, and returns the shared meeting entity/list envelope.

meeting_project_evidence accepts an existing project name/key/alias and the shared selection query plus limit, cursor and max_read_bytes. It returns one registered item with project, matched name, cited hits, coverage and unconfirmed association status. Each call reads one name page from the existing index without refreshing it. Follow the root next step even after an empty hit page; returned dates are frozen for replay. It never creates aliases or confirms project membership.

After installing the skill, start a new agent session if it has not discovered the instructions. After adding MCP, restart the AI app and confirm that its Zoom tools are available. Ask:

List my stored meetings. Open the latest one and show its date, available transcript and one
passage supporting its main decision. Tell me if the data is incomplete.

Check the selected meeting against zm meetings list --json in your terminal. If discovery works but no data appears, verify the same profile and store path, then pull or import data. If the app cannot launch the command, use where.exe zm on Windows or command -v zm on Linux/macOS to find it, and put that executable's full path in the entry. A remote web chat cannot run this local command unless it has a supported connection to your machine.

Successful tool replies use the same item/list envelopes as the CLI, with snake_case keys and meta.run_id / meta.duration_ms. Detailed context, participant and semantic reports are inside item; standard meeting, transcript, summary and word-hit lists use items and page. Text content and structured content contain the same complete answer. Every tool advertises a strict success/error output schema. Unknown arguments are rejected in the handler before a store opens, with validation_error naming the field. The server selects the account.

Optional flags add tools without changing the defaults: --meeting-context adds cited passages, evidence, person context and task proposals; --meeting-preparation adds meeting_prepare for remote schedule plus cited stored context; --unified-search adds meetings-only shared search; --semantic --model <model> lets search_meetings rank existing vectors by meaning (by_meaning). Remote semantic models additionally require --base-url and --dims. The server fixes the model at startup and the profile on its first read; restart it to change either. Clients cannot override either selection. Embedding generation and task creation are never MCP tools.

For broader meeting context across projects, notes and documents, continue with search and evidence; connect Memo separately where needed.

With zm serve --mcp --meeting-preparation, the agent can call meeting_prepare with series_id and optional ISO after, until, history_after, plus meetings, cues, context_limit, max_read_bytes and bytes. The server selects the account and reads Zoom schedule details, requiring meeting:read:meeting. This opt-in tool is unavailable with --offline; it does not change your Zoom app, retrieve absent transcripts or accept findings. Its meeting_preparation item retains cited cues, bounded related meetings through explicit person/confirmed project links, and explicit coverage. Accepted person/project facts appear in knowledge, with fact references and explicit unqueried-source/evidence limits. Owner notes and annotations appear separately in notes, retaining note: references, revisions and source/coverage limits. Imported documents, mail and messages remain unqueried. Related reads stay within the selected account and history window and perform no writes. See preparation for what it includes, what remains unqueried and how to check the result.

Preparation's bytes bounds the item after hidden characters are made visible, which can take more space than CLI JSON. Notes yield first, then accepted knowledge, related calls and exact history; inspect meta.truncated and each section's coverage for omissions. Cue content_bytes counts the escaped cues the MCP client receives. Metadata that cannot fit returns validation_error; increase bytes or narrow the selected context.