Zoom AI · Agentic Search · ZoomMate · MCP

Zoom Agentic Search: complete guide to enterprise AI search, ZoomMate, MCP, Search API, and Ask API

By Michael Obraitis, Managing Partner, and Michael Dale, CEO · Published August 28, 2026 · Last reviewed August 2026 · ADAM Pulse Knowledge Base

Verification date.

Product names, APIs, connector lists, licensing notes, and MCP setup steps were checked against Zoom and Microsoft documentation on 28 August 2026. These details change. Confirm against live vendor pages and your own tenant before you buy, connect, or process regulated data.

Short answer

Zoom Agentic Search is a unified context layer. It combines Zoom meetings, chat, phone, and notes (what Zoom calls dynamic context) with documents and records from connected enterprise systems (static knowledge). That layer serves ZoomMate inside Zoom Workplace and external agents such as Claude, Codex, OpenAI, or a custom MCP client. It is not just another search bar.

Architecture, as Zoom describes it on 11 August 2026: sources → Agentic Search context layer → Search API or Ask API → agents. Search API returns ranked evidence for your model. Ask API has Zoom synthesize a grounded answer on the server. MCP exposes the layer. Permissions are supposed to stay those of the source, via identity mapping on each connector.

Unlocking the full layer, per Zoom, takes a ZoomMate plan for Ask API and for the ZoomMate surface; Search API is documented against an eligible Zoom Workplace or ZoomMate paid plan. Confirm current packaging on Zoom's order pages. This guide is the cluster pillar. Definition, API choice, MCP setup, Copilot Search comparison, ZoomMate versus classic enterprise search, and connector how-tos live in the companion pages.

What is Zoom Agentic Search?

Zoom's research post (Ben Tezcan, 11 August 2026) frames the problem this way: a context layer that only indexes static knowledge — wikis, files, CRM records already written down — misses where most work actually happens. A deal's real status may exist only in this morning's call. Who is carrying this week's launch may exist only in standup. Zoom calls that dynamic context.

Agentic Search is Zoom's attempt to unify both: static content from connected third-party platforms plus the live and recent conversation that lives in Zoom. An agent asks one question and is supposed to receive both the settled document and the current conversation, ranked and ready to reason over.

Two access paths, same layer:

Two ways Zoom documents access to Agentic Search.
Path	Who it is for	What you get
ZoomMate	People already in Zoom Workplace	Ask in ZoomMate; Zoom says meetings, docs, chat, and 10+ connected sources sit behind the question.
MCP	Teams that want Claude, Codex, OpenAI, or a custom UI	The same layer over Zoom's MCP server, via Search API or Ask API.

That is different from keyword search over a file share, and different from a chatbot that only sees the current thread. For the one-sentence definition, see What is Zoom Agentic Search?

How does the architecture work?

Keep this picture, then map your tenant onto it:

Sources. Zoom-native content (meetings, Team Chat, phone, Canvas / My Notes, recordings and summaries where enabled) plus third-party systems connected in AI Studio as data sources, and separately ZoomMate user-level connectors for actions.
Context layer. Agentic Search ranks and retrieves across those sources for the signed-in user.
APIs. Search API returns ranked results. Ask API returns a synthesized, cited answer after a server-side retrieval loop.
Agents. ZoomMate, or any MCP-capable client that Zoom's servers support.

Zoom's pitch against federated, tool-by-tool retrieval is that an agent which calls each connector itself spends tokens on plumbing: more tool calls, more reasoning loops, more half-relevant payload in the context window. A ranked layer is supposed to return the relevant material directly. Whether that holds in your tenant depends on source quality, identity mapping, and which API you call. It is not automatic.

When should you use Search API versus Ask API?

Zoom is explicit that “context layer” means something different depending on who is building:

Search API — ranked results for teams that want to bring their own model. Evidence in, your synthesis out.
Ask API — retrieval and synthesis both happen server-side. The agent sees one request and one grounded answer. Intermediate transcripts and fan-out do not land in the agent's context window.

Licensing footnotes from the same Zoom post: Search API requires an eligible Zoom Workplace or ZoomMate paid plan; Ask API requires a ZoomMate paid plan and available Zoom AI credits. Token figures in that post include both Claude and ZoomMate token usage where each is involved.

The decision table lives in Zoom Search API vs Ask API. Short version: use Search API when you must inspect citations, apply your own policy, or keep synthesis in a model you already govern. Use Ask API when the job is a grounded answer and you accept Zoom doing the retrieval loop.

How does MCP expose the layer to Claude, Codex, and other agents?

Model Context Protocol is the open standard Zoom uses so an external agent can discover tools and retrieve Zoom context without a custom integration per client. Zoom documents this in two complementary ways, and you should not mix them up:

Published connectors / plugins. Zoom for Claude Connector (Claude's connector directory) and Zoom for Codex plugin. These are the shortest path for those two products.
Remote MCP servers. Product-specific servers (Meetings, Chat, Canvas, Tasks, Whiteboard, Revenue Accelerator) plus a unified Zoom MCP surface used by Zoom's own Claude plugin. Zoom's developer overview gives Streamable HTTP and SSE examples such as https://zoom.us/mcp/meeting/streamable. Zoom's plugin catalog also documents https://mcp.zoom.us/mcp/zoom/streamable. Copy the URL from the current server page for the product you are connecting. Do not hard-code a blog example.

Remote MCP requires a Zoom App Marketplace General app, OAuth scopes taken from that server's tool list, and a Bearer token. Zoom does not support Dynamic Client Registration or Client ID Metadata Documents on these servers. MCP clients must send PKCE (code_verifier) even though Zoom's token endpoint is classic OAuth 2.0. Setup steps: Connect Claude AI agents to Zoom MCP.

Zoom's Claude connector docs also add a practical limit: meeting assets and recordings are available when Smart Recording and/or Meeting Summary are enabled and you hosted the meeting or the host granted access. Video features need cloud recording with Smart Recording. If those admin toggles are off, MCP search looks empty. That is a configuration gap, not a broken agent.

What does Zoom's 60 to 80 percent token figure actually measure?

Do not quote it as a guarantee. Quote the test.

Zoom ran the same request — “Review the last week of meetings and compile the follow-ups” — through Claude twice. First run: raw retrieval only (list meetings and open them; brute force). Second run: Claude connected via MCP and used Search API. Same agent, same week, retrieval path as the only variable.

Zoom then states 60% to 80% fewer tokens for “one ranked call instead of eleven brute-force reads.” Methodology on the same page: figures are directional, drawn from a real run over a single user's 7-day calendar (89 meeting instances), rounded for illustration. The Ask API figure is a projection at roughly half of Search API retrieval, pending a direct benchmark. Actual token counts vary by query, payload size, and connected sources, and exclude fixed system-prompt overhead.

What that supports: ranked retrieval can beat brute-force open-every-meeting on that workload. What it does not support: a promise that your Salesforce-plus-SharePoint-plus-ZoomMate deployment will cut tokens by 60 to 80 percent, or that Ask API has been benchmarked the same way. Measure your own pilot.

How are permissions preserved?

Zoom's AI Studio data-source articles (SharePoint, Salesforce, ServiceNow, Google Drive, and the umbrella KB0084343) repeat the same control: account mapping between Zoom users and source identities. Zoom says it accesses content on behalf of the mapped user and adheres to the source's permission structures. Partial mapping is expected. Unmapped users are not included.

That is the claim to design around. It is not a substitute for a source-side access review. If SharePoint permissions are already a mess, Agentic Search will faithfully retrieve the mess for whoever is mapped. Restricted SharePoint Search and similar Microsoft controls still matter on the Microsoft side; Zoom mapping does not replace them.

Sync cadence from KB0084343 (verify in tenant): OneDrive, SharePoint, and Google Drive full sync every 28 days with incremental sync about every 10 minutes; other apps full sync daily with no incremental. You can create only one app instance for OneDrive / SharePoint / Google Drive; other apps can have multiple instances. Connector how-to: Connect Salesforce, ServiceNow, SharePoint and Google Drive to Zoom AI.

How should you deploy it in phases?

Treat Agentic Search as a managed capability, not a connector checkbox. A sequence that matches how we already tell customers to roll out Zoom AI:

Phase 1: Discover

Inventory Zoom Workplace plan, ZoomMate seats, AI credits, regions, and app versions.
List candidate sources. Separate knowledge (AI Studio data sources) from actions (ZoomMate third-party apps).
Classify data. Decide what must never enter an agent context window.

Phase 2: Design

Pick two or three measurable questions the layer must answer (not “turn on search”).
Choose Search API versus Ask API per use case — they are not interchangeable.
Plan identity mapping and an admin who owns incomplete maps.
Write the human-review rule: what an agent may draft, what a person must send.

Phase 3: Pilot

One group, one or two sources, questions you can score.
Log empty results, wrong citations, and permission surprises separately.
If you connect Claude or Codex, confirm OAuth scopes and that Meeting Summary / Smart Recording are actually on.

Phase 4: Launch

Expand sources only after mapping and quality hold.
Keep ZoomMate (worker surface) and MCP (developer / agent surface) in the same governance review.
Re-check Zoom's connector catalog. The list grows; do not freeze a blog post as the catalog.

ADAM Pulse can help design that pilot. We do not flip production connectors without a named owner and a rollback.

What should you not expect?

Claims this cluster will not make.
Expectation	What to do instead
A guaranteed token or license saving	Use Zoom's published test as a directional illustration. Measure the pilot.
A replacement for Microsoft 365 Copilot Search	Different grounding (Graph + Copilot connectors versus Zoom layer + AI Studio sources). Many tenants will run both. See the comparison companion.
Instant completeness on day one	Sync windows, mapping gaps, and disabled Meeting Summary all produce empty or stale answers.
Ask API output as the system of record	Keep CRM, ITSM, and file platforms authoritative. Require review before anything leaves the team.
Every Zoom MCP server exposing Search and Ask	Product servers expose product tools. Agentic Search APIs are the context-layer path Zoom describes on the research post and MCP surface. Confirm tools with a live `tools/list`.

How does this relate to the existing Zoom AI Companion articles?

Those pages still answer different questions. This cluster does not replace them.

Zoom AI Companion in 2026 — what is included, what ZoomMate adds, how to deploy the assistant features.
Is Zoom AI Companion free? — packaging and ZoomMate pricing.
Zoom AI vs Microsoft 365 Copilot — licensing the two assistants. The new comparison is Agentic Search versus Copilot Search, a different Microsoft surface.
Does Zoom use your meetings to train AI? — the no-training terms. Connector data is an extra governance review, not a loophole in that article.

Frequently asked questions

Is Zoom Agentic Search just another search bar?

No. Zoom describes it as a unified context layer: meetings, chat, phone, notes, and connected enterprise systems, served to ZoomMate and to external agents over MCP. A search box is one surface. The layer underneath is the product.

What is the architecture?

Sources feed the Agentic Search context layer. Agents then call Search API (ranked evidence for your model) or Ask API (Zoom synthesizes a grounded answer). MCP exposes those APIs to Claude, Codex, OpenAI, and other MCP clients. Permissions stay those of the source.

What is the difference between Search API and Ask API?

Search API returns ranked results so your model can reason over evidence. Ask API runs retrieval and synthesis on Zoom's side and returns one grounded answer. Zoom says Search API needs an eligible Zoom Workplace or ZoomMate paid plan; Ask API needs ZoomMate plus available Zoom AI credits.

Does Zoom guarantee 60 to 80 percent token savings?

No. Zoom published 60 to 80 percent fewer tokens from a specific Claude test: brute-force list-and-open versus Search API, over one user's 7-day calendar with 89 meeting instances. The Ask API figure is a projection pending a direct benchmark. Treat it as Zoom's directional test, not a guarantee for your tenant.

Does Agentic Search bypass source permissions?

Zoom's AI Studio data-source docs say identity mapping is required so ZoomMate retrieves only content the user can already see in the source. Partial mapping means unmapped users do not get that source. Do not assume a connector copies the whole repository to every Zoom user.

How do Claude or Codex get this context?

Through Zoom's MCP servers. Zoom publishes a Zoom for Claude connector in Claude's connector directory, custom-connector setup for Claude.ai, ChatGPT Developer Mode setup, and product-specific MCP servers (Meetings, Chat, Canvas, Tasks, Whiteboard, Revenue Accelerator). Remote MCP requires a Zoom Marketplace General app and OAuth with PKCE. There is no Dynamic Client Registration.

Is this the same as Zoom AI Companion?

No. Core meeting, chat, and phone AI features that used to carry the AI Companion name remain included with eligible paid Zoom Workplace plans. ZoomMate is the agentic add-on. Agentic Search is the context layer ZoomMate and MCP agents use. See the existing Zoom AI Companion guides linked below.

Should we put regulated or confidential data into the layer on day one?

No. Pilot on a classified, low-risk source set. Confirm identity mapping, retention, and human review before you connect systems of record. Do not treat a synthesized Ask API answer as the official record.

This page is part of the ADAM Pulse Zoom Agentic Search cluster. Existing Zoom AI Companion guides are related, not duplicates.

References

Primary vendor documentation. Where a vendor has not published a claim, this article says so rather than filling the gap.

ADAM Pulse can help design an Agentic Search pilot

ADAM Pulse helps Zoom customers pick sources, map identities, choose Search API versus Ask API, connect Claude or other MCP clients, and run a phased pilot with human review. This is a design engagement, not a product claim about Zoom. Support: (888) 989-4872 · support@adampulse.us

Prepared by ADAM Pulse (USA Telecom Consulting LLC) — managed Zoom and network services. Support: (888) 989-4872 · support@adampulse.us