What can MCP and computer use actually do for event ops—and what still needs a person?
By Donghui Qian (Milo)Updated 17 November 20267 min read
Short answer
MCP lets a connected AI call authorised event tools—create unpublished drafts, query channels and inventory, pull summaries for a retrospective. Computer use helps the same person check what attendees actually see in a browser. Neither replaces organiser review: prices, conflicts and publish stay human. Measure hours saved on drafting and lookup, not a mythical tenfold leap across the whole event.
MCP (Model Context Protocol) lets a connected AI client call authorised tools on your event stack so drafting and lookups get faster: unpublished site and form drafts from existing materials, speaker and agenda sheets mapped into structure, channel and inventory checks before you send. Computer use is how that same person walks preview and live pages in a browser to see what attendees actually see. Used together—under clear authorisation—they save hours on those jobs. They do not run the event alone.
This guide is for organisers and ops leads evaluating AI connectors on a real stack. Product rules for bagevent’s in-app assistant stay on the AI assistant page; the lean-team AI layers guide decides what belongs in chat versus the registration record. Here the question is narrower: what MCP and computer use are for, with examples from our UC project work, and what must stay a human gate.
In brief: two tools, one human gate
- MCP: authorised tool calls against event data and draft actions.
- Computer use: look at the public page, mobile layout, and what attendees see.
- AI clients may also read files (Excel, design refs)—that is not the same as an MCP write.
- Drafts stay unpublished until an organiser checks prices, dates and conflicts.
- Score time saved on concrete jobs (first draft, field mapping, channel lookup)—not “10× the whole week”.
MCP versus computer use—do not merge the story
| Capability | What it is for | What it is not |
|---|---|---|
| MCP tools | Create or update structured drafts; query authorised events, tickets, channels, check-in aggregates | A guarantee that every client can write; Pro-tier or plan limits may be read-only—confirm in the client |
| File / vision in the AI client | Parse speaker sheets, extract photos, propose field maps before anything is written | A dedicated “upload Excel into MCP” importer |
| Computer use (browser) | Open the public or preview URL; check banner, nav, mobile; spot missing photos | A substitute for MCP data queries or for clicking Publish without a person |
If a write-up collapses these three into “the AI built the event,” readers cannot tell which step was a tool call, which was file parsing, and which was a human in the CMS. Our UC notes keep those labels separate on purpose.
Connect once, then name the event every time
You need the full MCP service URL from the vendor—not the marketing homepage or a registration page—and an account that can see the target events. In ChatGPT, custom MCP is added under Plugins (or the current Apps / connectors UI); Claude uses Connectors. Team workspaces often need an admin to add the connector first. Account plans change: as of late 2026, OpenAI documents fuller MCP for Business and Enterprise/Edu, with tighter write rights on some Pro setups—always re-check the vendor help page before you promise a workflow to your team.
Check the bagevent connection: how many events am I authorised for,
can I create new drafts, and list the events I can access.
In our verification pass the reply named two authorised events and confirmed draft creation. If the expected event is missing, fix account and event grants before blaming the model. After connect, every request should name the event, the data needed, and the output shape—and, for new drafts, which facts are already confirmed.
Prep: site draft from brief, old site, and design—not a pixel clone
When you already have a previous public site, a design board, and a content Word file, hand all three to the connected client. Ask it to list pages, reusable assets, and open questions; keep the visual language; verify desktop and mobile; do not carry over unverified dates or fees.
UC example (CCMP). Session logs show the model called the MCP draft-create tool. The first save held on the order of seventeen pages, a committee roster, and a registration form, reusing banner, logo and photo assets from the prior site. Later passes adjusted banner, secondary nav, people cards and footer—some in the browser editor, some in product engineering. The public screenshots people see are after iteration, not the first MCP write. MCP saved an unpublished draft; organisers still review and publish. This is a head start from existing materials, not one-click pixel replication.
Prep: speakers and agenda from a sheet
Large programmes lose hours copying name, title, company, bio, headshot, session time and speaker links. Prefer: upload the sheet to a client that can read files → list field maps, duplicates, gaps and time conflicts → confirm → then write into a new draft via the tools that actually accept speaker and session payloads.
- UC example (Beijing developer event). An updated speaker Excel was read in-session; four photos in one row were matched to the roster; gaps were filled only after confirmation; sessions were linked in sheet order. When a workshop title listed two speakers but the bio column had one, the correct move was to flag the mismatch—not delete a name or invent a bio.
- Split of labour. File read and vision stay in the AI client. MCP draft APIs accept structured speaker/agenda payloads; they are not a magic Excel uploader. Updates to an existing live event may still need backend or other authorised paths—do not narrate every fix as “MCP wrote it.”
Check: records in the system versus the page people see
A row in the admin is not proof the public page shows the right crop, the right session link, or the mobile footer. This is where computer use earns its keep: open preview and live URLs, walk the phone width, list differences against the confirmed brief. MCP queries can supply facts for that checklist; they do not replace eyes on the page.
Series events (for example Beijing and Shenzhen sharing a speaker library) need an extra rule: what belongs to the org library versus this city’s event. Sync shared assets, then re-check each city’s dates, venue, agenda and registration entry separately.
Operate: channels, inventory, multi-city briefs
- Partner entry. Query the invite channels on an event, pull the existing registration link and share image for a named channel, draft forward copy—then humans approve and send. Current tools read existing channels; they do not invent new ones. The share image is the event creative, not a channel-specific QR poster unless your stack says otherwise.
- Inventory before a blast. Pull ticket totals, sold and remaining with price at query time. Draft a reminder that does not invent “almost sold out.” Re-query before the real send.
- Multi-city week. Pull orders, registrants, check-in aggregates and revenue per city as separate columns. Orders ≠ registrant headcount—show both. Same-day check-in totals are a snapshot, not a live CCTV feed, and aggregate tools must not dump attendee PII into the chat.
Close: retrospective that separates facts from guesses
Ask for a draft retro that splits confirmed system facts, explanations still to verify, and next actions—with a hole list for missing data. Channel volume is not ROI without cost and downstream outcomes. Confirmed lessons become the next brief’s inputs; the organiser still owns the decisions.
Where hours actually move—and where they do not
| Job | Often faster with MCP / computer use | Still human |
|---|---|---|
| First site + form draft from old materials | Structure, asset reuse, unpublished draft | Fees, legal copy, brand sign-off, Publish |
| Speaker / agenda sheet into a new draft | Mapping, gap lists, bulk structured write after confirm | Resolving roster conflicts; live-event patches |
| Channel link + promo draft | Find channel, pull link/image, draft text | Partner relationship, send time, claims |
| Public page QA | Browser pass on desktop/mobile against a checklist | Taste, accessibility edge cases, go-live |
| Door and money truth | Aggregate queries when authorised | Offline door ops, refunds, settlement |
That table is the honest substitute for “10× efficiency.” Some tasks drop from hours to a reviewed draft; the week still has doors, partners and finance. If a vendor demo only shows the chat and never the unpublished-draft → human Publish path, you are still shopping on theatre.
Frequently asked questions
MCP (Model Context Protocol) lets a compatible AI client call tools the event platform exposes under your login—such as creating unpublished drafts or querying ticket and channel data. It is not a guarantee of write access on every AI plan, and it is not the same as the model reading a file in chat.
Computer use drives a browser to inspect preview or public pages. MCP calls structured tools and data. Use computer use to check what attendees see; use MCP to draft and query. Neither should publish without an organiser.
It can create a strong unpublished draft from a prior site, design refs and a content doc—as in our CCMP UC work—then humans iterate layout and facts. It is not a promise of pixel-perfect clone or automatic go-live.
No. The AI client parses the sheet; MCP draft APIs accept structured speaker and session data after you confirm the map. That is not a dedicated Excel importer, and live-event updates may use other authorised paths.
Do not buy that headline. Expect clear time saves on first drafts, field mapping, channel lookup and page QA—then keep human gates on prices, conflicts, publish, refunds and the door. Measure those jobs; ignore vague multipliers.
It proposes inside the product with confirm-before-write rules. MCP is how external clients such as ChatGPT or Claude call authorised tools. See the AI assistant feature page and the lean-team AI layers guide for the split.
Read next
- AI and event apps for lean teamsThree layers: general AI, event stack, in-product assistant
- AI assistantPropose → confirm; what the in-product assistant may draft
- Event website SEOWhat the public page must say—and llms.txt honesty
- Accept the shortlist after the showSame spirit: prove the door, not the demo theatre
- Multilingual registrationLanguage versions still need human checks after drafts
- Event websiteWhere drafts become the attendee-facing site