Zoom MCP Server: Setup, Tools & Real Limits
Zoom MCP now has first-party hosted servers, but useful results still depend on licensing, recording settings, narrow OAuth scopes, and disciplined result retrieval.
Zoom MCP is Zoom’s official bridge from an MCP-compatible AI client to Zoom Workplace data and actions. It can search meetings, retrieve transcripts, summaries and recordings, inspect chat and documents, and create selected follow-up content. Use the hosted first-party server for supported workflows, but start read-only: its usefulness is real, and so is its access to sensitive conversations.
Table of contents
- What Zoom MCP actually is
- What exists today
- How to set up Zoom MCP
- Tools you get
- Permissions and blast radius
- Real limits
- When not to use it
- Our verdict
What Zoom MCP actually is
Zoom MCP is a remote tool layer, not a meeting bot or AI model. A client discovers operations from a Zoom-hosted endpoint, calls them under an OAuth identity, and gives returned data to the model for reasoning or drafting.
An agent can locate where a decision was made, fetch the transcript and summary, return a cloud-recording link, search Team Chat or Zoom Docs, and create a follow-up Doc from Markdown. Zoom’s official MCP overview explains the client–server model; its registry identifies the hosted endpoints. The MCP Directory capabilities index separates search, retrieval and write actions.
MCP does not grant new Zoom rights. Zoom still applies the signed-in user’s licenses, scopes and content access, so “connect Zoom” does not mean “search every meeting in the company.”
What exists today
Zoom now has first-party hosted servers, making claims that every Zoom MCP server is community-maintained obsolete. Community and managed alternatives remain separate implementations with different tools, authentication and support owners.
Zoom’s registry lists Workspace, Meetings, Team Chat, Docs, Whiteboard, Tasks and Revenue Accelerator servers. Workspace covers several products; Meetings is the better default for transcript and recording retrieval. Both use streamable HTTP. See the Zoom MCP registry and MCP Directory’s official collection.
| Option | Best fit | Trade-off |
|---|---|---|
| Zoom Workspace MCP | Cross-product search | Broader data surface |
| Zoom Meetings MCP | Meetings, transcripts, recordings | Narrower product coverage |
| Zoom for Claude | Fast Claude user setup | Plan and AI Companion dependencies |
| Workato Zoom Meetings MCP | Governed meeting CRUD | Paid intermediary |
| Community package | Niche or local experiments | You own security and maintenance |
We would choose the focused first-party Meetings server for production retrieval, adding Workspace only when a tested workflow genuinely needs chat and Docs together.
How to set up Zoom MCP
Configure the hosted server as a remote OAuth connector; do not install an unverified package because an old guide shows npx. Zoom requires a Marketplace app, selected scopes, OAuth and the product-specific URL.
- Confirm the product licenses, MCP-capable client, and Zoom admin or developer authorization. Recording workflows need the relevant recording and AI Companion features.
- In Zoom App Marketplace, choose Develop → Build app → General app.
- Add only required scopes. Begin with meeting, transcript or recording reads; postpone create, update, chat-send, Doc-create and task-write permissions.
- Add the registry endpoint as a remote connector. Meetings uses
https://mcp.zoom.us/mcp/meeting/streamable; Workspace useshttps://mcp.zoom.us/mcp/zoom/streamable. - Supply the app’s client ID and secret, register the client’s redirect URL, complete OAuth, and test one read.
Zoom’s connection guide says remote servers require manual registration, do not support DCR or CIMD, and require PKCE. We omit a generic mcpServers block because remote OAuth configuration varies by client; use the config generator after selecting the client and registry URL.
Tools you get
Search and retrieval form the core tool surface; writes depend on the selected server. Machine identifiers are discovered from that server, so the table separates Zoom’s verified operation labels from exact names published for Workato’s different implementation.
| Operation or tool name | Distribution | What it does |
|---|---|---|
| Search meetings | Zoom Meetings | Finds meetings by keyword and date |
| Review meeting assets | Zoom Meetings | Gets summaries, transcripts and recordings |
| Inspect recordings | Zoom Meetings | Lists cloud recordings and resources |
| Create follow-up Docs | Zoom Workspace | Creates a Doc from Markdown |
search_meetings | Workato | Searches hosted or attended meetings |
get_meeting_transcript | Workato | Gets an auto-generated transcript |
get_meeting_participants | Workato | Returns attendance |
create_meeting / update_meeting | Workato | Creates or changes a meeting |
Zoom’s labels come from its Meetings registry metadata, not guessed identifiers. Workato publishes its exact tool list. Do not assume Workato’s names exist on Zoom’s endpoint; inspect the discovered schemas.
Permissions and blast radius
Treat meeting content as confidential and authorize reads before writes. A careless write can move a customer meeting, expose a join link, send a false chat message, or create misleading follow-up material.
For unattended workflows, use a dedicated integration user where permitted; otherwise use individual OAuth so callers retain their own boundary. Never authenticate a general-purpose agent as the account owner. Admin scopes can turn one bad prompt or compromised client into account-wide disclosure.
Our rollout sequence is:
- enable Meetings, not Workspace;
- grant only required read scopes;
- confirm before returning shareable recording links;
- log calls without transcript bodies or tokens;
- add writes individually, with confirmation and a staging user.
Zoom’s Claude connector documentation states that users can retrieve assets only when they hosted the meeting or received host access. That boundary should remain intact.
Real limits
The real limits are source availability, pagination, context size and API quotas. The server cannot fetch a transcript Zoom never produced, and a model cannot reason reliably over an unlimited meeting dump.
Recording intelligence depends on meeting settings. Zoom says video features require a Smart Recording in the cloud, while asset access may be limited without Smart Recording or Meeting Summary. Local recordings are not equivalent, and Zoom’s API says end-to-end encrypted meetings do not support AI Companion summaries.
Retrieve narrowly. Zoom APIs paginate collections; one meeting listing defaults to 30 records, and next-page tokens can expire after 15 minutes. Long VTT transcripts consume context quickly, so filter by date and topic, fetch metadata, then retrieve one transcript or segment. The Zoom Meetings API reference documents pagination and prerequisites.
Quotas vary by endpoint and plan. A limit breach returns HTTP 429; meeting creation is documented at 100 requests daily per host, while past-participant data has a 15-month window. Use bounded retries, cache stable metadata, and avoid account-wide fan-out.
MCP also does not stream live media. For that, Zoom’s RTMS starter kit receives active-meeting transcript, audio, video and screen-share streams—a different architecture and consent burden.
When not to use it
Skip Zoom MCP when Zoom’s own AI features answer the question or a fixed API job is easier to test. MCP fits varied, human-driven questions; it adds unnecessary indirection to deterministic automation.
Use AI Companion or a supported connector for recap, action items or finding a moment in one recording. The Zoom for Claude documentation already supports meeting search, summaries, playback links and assets.
Choose a plain API script for nightly exports, retention, attendance sync, or form-driven scheduling. It has a smaller prompt-injection surface and clearer tests. Choose RTMS for live media. Avoid MCP if policy forbids sending transcript content to the model provider.
Our verdict
Zoom MCP is worth using for human-driven meeting retrieval, but launch with a narrow first-party server and read-only scopes. The trade-off is flexibility versus control: natural-language search saves time, while every extra product and write action increases privacy and failure costs.
We would cap searches by date and result count and keep writes behind confirmation. We would not bulk-feed transcripts or base critical scheduled work on agent judgment. Browse MCP Directory’s best MCP servers, but judge Zoom integrations by publisher, scopes and discovered schemas.