MCP (Model Context Protocol) lets an AI assistant use another product’s tools. AtriumKit runs an MCP server, so you can ask Claude, ChatGPT or Cursor to fix a renamed label, rewrite tooltips, translate a demo or start a video without opening the editor. This tutorial gives you the setup for each client, prompts to paste, and what each prompt changes, costs and leaves for you to do.
We make AtriumKit, so weigh what we say about it accordingly. We’re also not the only demo tool with a server: Storylane, Supademo, Navattic and Arcade have them too.
Before you start
- An AtriumKit account with a recorded demo. The MCP server works on every plan, Free included. If you don’t have a demo yet, record your first one.
- The editor or admin role, if you want to change anything. A viewer’s assistant can read but not edit.
- The server URL, which is the same for every client. It uses Streamable HTTP, and clients sign in with OAuth 2.1 or send a personal access token.
https://app.atriumkit.com/mcp
Step 1: Connect your assistant
Client menus change often. We checked these steps against each vendor’s docs on November 5, 2026, and linked the docs so you can check them again.
| Client | Where you add it | Plan |
|---|---|---|
| Claude | Customize > Connectors | Every plan. Free can add one custom connector. On Team and Enterprise, an Owner adds it |
| ChatGPT | Plugins, on the web | Business, Enterprise and Edu get read and write. Pro gets read and fetch only, so edits won’t work. Plus and Free aren’t listed |
| Cursor | mcp.json |
Cursor’s MCP docs list no plan requirement |
Claude
On Free, Pro, Max or Ultra (Anthropic’s docs):
- Go to Customize > Connectors and click Add custom connector. On some accounts it’s under + Add.
- Name it AtriumKit and paste the server URL. If you see Continue, click it.
- If the dialog asks about sign-in, choose Sign in now. Under OAuth client, choose Use Claude’s published identity. If you see client ID and secret fields instead, leave them empty.
- Click Add. If Claude shows Connect, click it.
On Team or Enterprise, an Owner goes to Organization settings > Connectors, selects Add, then Custom (and Web if asked), enters the URL and clicks Add. Each member then finds it under Customize > Connectors, with the Custom label, and clicks Connect.
To use it in a chat, click +, select Connectors and turn AtriumKit on.
Claude Code takes one command. Then run /mcp in Claude Code and follow the sign-in in your browser (docs). Add --scope user to use it in all your projects.
claude mcp add --transport http atriumkit https://app.atriumkit.com/mcp
ChatGPT
Check your plan first. OpenAI’s help article puts full custom MCP support, with write actions, on ChatGPT on the web for Business and Enterprise/Edu. It says Pro accounts can connect custom servers for read and fetch only, which rules out edits and publishing. Plus and Free aren’t listed, and it doesn’t work on mobile.
You don’t need to turn on developer mode. You accept a risk warning for each server instead.
The steps, from OpenAI’s docs:
- In ChatGPT on the web, open Plugins, select the plus button, then Add custom MCP server. On Business, admins and owners do this. On Enterprise and Edu, admins can let others do it.
- Name it AtriumKit and enter the server URL under Connection.
- For authentication, choose OAuth. If you see client ID and secret fields, leave them empty.
- Read the risk warning, select I understand and want to continue, then Create as a plugin.
- Install the plugin, start a new conversation, type
@and select it. Sign in to AtriumKit when ChatGPT asks.
Cursor
Add this to ~/.cursor/mcp.json (every project) or .cursor/mcp.json (one project):
{
"mcpServers": {
"atriumkit": { "url": "https://app.atriumkit.com/mcp" }
}
}
Cursor signs in with OAuth when the server asks for it (Cursor’s docs). To turn the server on or off, open Customize in the sidebar and find it.
Or use an access token
For scripts and clients that can’t sign in:
- In AtriumKit, open Settings (click ⋯ next to your name at the bottom of the sidebar, then Account settings) and choose AI assistants.
- Under Access tokens, click New token.
- Give it a Name, choose What it can do, pick when it Expires (30 days, 90 days, 1 year or No expiry) and enter Your password.
- Click Create token and copy it. It’s shown once.
The client sends it as Authorization: Bearer <token>. In Cursor, keep it in an environment variable rather than in a shared project file:
{
"mcpServers": {
"atriumkit": {
"url": "https://app.atriumkit.com/mcp",
"headers": { "Authorization": "Bearer ${env:ATRIUMKIT_TOKEN}" }
}
}
}
Approve it on the AtriumKit page
With OAuth, your client sends you to an AtriumKit page before anything connects.
- Check the address. The page marks every app Unverified app, because an app can call itself anything, and it shows where your answer goes back to. Claude Code and desktop Cursor return to
localhost, and the page warns that any program on your computer can start such a request. Continue only if you just connected it yourself. - Pick the workspace. Each connection works in one workspace.
- Pick what it may do. Read demos is always included. The others are Edit drafts, Publish, Read analytics and Read leads.
- Click Allow access.
You shouldn’t have to repeat this often. AtriumKit issues refresh tokens, so clients that support them renew their one-hour access on their own. A connection lasts 90 days from its last use. After that, or after you change your password, connect again.
Then test the connection:
Which AtriumKit workspace are you connected to? Tell me my role, the plan and how many AI credits are left.
The assistant calls list_workspaces. Nothing changes.
Step 2: Try these prompts
The demo names are examples, so use your own. The names in code font are the tools the assistant should call. You don’t have to name them, though naming one helps if the assistant picks a different route.
1. See what’s published
List every published demo in this workspace with its title, link and views. Tell me which ones have unpublished changes.
- Tool:
list_demos - Changes: nothing. It’s read-only.
- Cost: no AI credits.
The answer also gives the assistant the ids it needs for the next prompts.
2. Fix a renamed label or a price
Start with a dry run, which saves nothing:
We renamed “Members” to “People”. In the demo “Invite your team”, do a dry run of replacing “Members” with “People” in the copy, whole words only and matching case. List each place it would change.
Then:
That looks right. Apply it, then do the same for “members” to “people”. Don’t publish.
- Tool:
replace_in_copy - Changes: step titles, bodies, button and choice labels, hotspot labels and hints, the description, variable defaults and guide overrides, in one draft save.
{{variables}}are left alone. Narration scripts change only if you ask, and their clips then need generating again. - Goes live: no. It stays in the draft until someone publishes.
- Cost: no AI credits, on every plan.
Those two options matter. Matching ignores case by default and inserts exactly what you typed, so “members” would become “People”. Without whole words, “Membership” becomes “Peoplehip”, and $49 also matches inside $490.
Captured screens still show the old word. An assistant can’t change text on them, but it can find it:
Look at each screen of “Invite your team” and tell me which ones still show “Members”.
get_screen returns each screen as viewers see it, with its readable text. Then open the demo in the editor. On an HTML screen, choose Edit text and click Find & replace… to change the word on every HTML screen at once. On a screenshot, use Edit text. Editing captured screens has the details.
When you’re done, click Publish changes. If the demo has translations, the changed strings are now out of date, so run the translate prompt below too.
3. Rewrite tooltips for one audience
Duplicate “Send an invoice”. In the copy, rewrite every tooltip for a CFO: shorter, and mention month-end close. Keep every
{{variable}}. Show me the before and after for each step. Don’t publish.
- Tools:
duplicate_demo, thenrewrite_copy - Changes: makes a new draft named “Send an invoice (copy)” and saves the rewrite to it. The original doesn’t change.
- Goes live: no. The copy is a draft until someone publishes it.
- Cost: 1 AI credit per 20 strings, at least 1, so 30 strings cost 2 credits.
rewrite_copy saves by default. Say “suggestions only” to review the text first, which still uses credits. To spend no credits at all, ask the assistant to write the text itself and save it with update_step (guided demos) or update_content (hotspots in clickable demos).
4. Translate a demo and publish it
Translate “Send an invoice” into German. Tell me how many strings you translated, how many credits that used and whether anything needs review. If nothing needs review, publish it.
- Tools:
translate_demo, thenpublish_demo - Changes: adds the language and translates missing and out-of-date strings.
- Goes live: only at
publish_demo, and only if the connection has Publish. Without it, you click Publish changes yourself. Publishing puts the whole draft live, so any other unpublished edits go with it. - Cost: 1 AI credit per 50 strings, at least 1. AI translations are on Pro and Team.
A demo can be in 15 languages, English included: the one it’s written in and up to 14 translations. Text inside captured screens isn’t translated, and narration audio for each language is generated in the editor.
5. Find the step that loses viewers
Which step of “Send an invoice” lost the most viewers in the last 30 days? Show me the funnel and suggest a clearer tooltip for that step. Don’t change anything yet.
- Tool:
get_analytics, which needs Read analytics. - Changes: nothing. If you like the suggestion, say “use it” and the assistant saves it to the draft with
update_step. - Cost: no AI credits. Free workspaces get view counts only. Step funnels are on Pro and Team.
Make a product demo video with Claude (beta)
Ask Claude for a video of a demo (ChatGPT and Cursor work the same way), and the server’s video tools make it in the video studio. You export the MP4 yourself. The studio is in beta, so expect it to change. You don’t need a separate “product demo video skill” for Claude or any other client, because the video tools come with the server.
Make a 45-second walkthrough video from the demo “Send an invoice”, 16:9, for finance teams. Tell me when it’s ready and give me the studio link.
- Tool:
generate_video, with the demo and thewalkthroughformat. The other formats areteaser(the default),launch,explainerandsocial. It takes up to 3 demos, 10 to 180 seconds, and 16:9, 9:16 or 1:1. - What it makes: a 1080p motion video built from your demo’s screens, with a script, music and narration, and stock footage or AI images where they help. It doesn’t record your screen. The AI director is told to show numbers and customer quotes only if you supply them.
- How long: usually 1 to 3 minutes. The tool waits 20 seconds, so the assistant may say it’s still generating. Ask again in a minute rather than asking it to start over, which would cost credits again.
- Goes live: no. Only your workspace can see the video until someone turns on Public link in the studio.
- Cost: 5 AI credits up to a minute, 10 for longer, refunded if it fails. AI narration needs Pro or Team and uses AI voice characters. On Free, the scripts stay in the scenes without a voice, and you need a verified email first.
Then edit it in the same chat:
In that video, change the first scene’s headline to “Send your first invoice”.
Rewrite the last scene so it ends on a “Start free” button.
Rename “Members” to “People” everywhere in the video.
update_scenechanges a scene’s text, timing, transition or narration script. It’s free.write_scenehas the AI director rewrite or add one scene for 1 AI credit.replace_video_textrenames across scene text, labels, narration scripts and the title. It’s free, but it leaves images alone, so screens in the video still show the old word. Changed scripts needgenerate_video_voiceovers, which uses voice characters.
You get the MP4 from the studio. Exports and share links aren’t available over MCP. Open the studio link, click Export, choose Final · 1080p, click Export MP4, then Download MP4. Free includes 2 minutes of exported video a month, with a “Made with AtriumKit” mark. Pro includes 60 minutes per creator, Team 3 hours per creator, and trials 10 minutes. Ask the assistant how much is left, and it calls get_video_usage.
What MCP can’t do
Nothing here watches your product. An assistant knows only what you tell it, and it can’t change what a captured page shows.
- It can’t record pages behind a sign-in or recapture screens. It can start a new demo from public web pages or screenshot links with
create_demo, but anything behind a sign-in is recorded with the Chrome extension. When a control moves or the layout changes, recapture the screen in the editor, with the Chrome extension, and its steps stay on what they sit on. A changed flow means recording it again; a clickable demo can take new pages with + Capture more screens. - It can’t edit captured screens. Text, blurs, hidden elements, images and crops on screens are edited in the editor, where Edit screens can also rewrite, restyle or rebrand them with AI (Pro and Team).
- It can’t export or share videos, delete demos or videos, or restore a version. Restoring happens under History.
- It can’t manage members, billing or integrations, or upload logos, images or fonts.
What it costs
| Free | Pro | Team | |
|---|---|---|---|
| MCP server | Yes | Yes | Yes |
| Published demos | 3 | Unlimited | Unlimited |
| AI credits a month, shared by the workspace | 25 | 500 | 2,000 |
| AI translations | No | Yes | Yes |
| AI voice characters a month, per creator | None | 60,000 | 120,000 |
get_analytics |
View counts | Step funnels, sources, sessions | Step funnels, sources, sessions |
| Exported video a month | 2 minutes, with the mark | 60 minutes per creator | 3 hours per creator |
| Version history | 7 days | 90 days | 1 year |
Each task above lists its own credit cost. Publishing a demo that’s already live doesn’t count against the limit of 3 again, and failed AI actions give their credits back. Each connection can make 120 requests a minute, and each person can run 12 AI actions a minute. See how AI credits work and pricing.
Check your access
What an assistant gets. It acts as you, in one workspace, with your role checked on every request. It can read your demos and drafts, and on Team your presenter notes. It sees screens the way viewers do, with blurs applied, and the text it reads has emails, phone numbers and long numbers masked.
Choose what to tick:
- Publish is the switch for human review. Untick it, and
publish_demo,unpublish_demoand changing who can view a demo all fail. With it, an assistant can also take a demo offline. - Edit drafts can spend your workspace’s AI credits and voice characters, and a few of its tools take effect without a publish: share settings in
update_demo, changes to a brand kit (published demos that use it change too),create_personalized_linksandadd_to_collection. - Read leads gives it names, emails and companies. Leave it off unless the task needs it.
Keep approvals on for tools that write. Captured pages can contain any text, including text written to mislead an AI. Cursor and ChatGPT ask before they use write tools by default. In Claude, review each approval request, and you can set a tool such as publish_demo to Blocked by selecting the connector under Customize > Connectors.
Check what it did. Edits show under History, marked like “Edits by Ana (via Claude)”. You can preview a version and click Restore this version, for as long as your plan keeps it. AtriumKit also emails you when an app connects or a token is created.
Revoke it. Open Settings and choose AI assistants, then click Disconnect next to an app or Revoke next to a token. It loses access right away. Each row shows when it was last used and from which country, and admins see every member’s. Changing your password or leaving the workspace disconnects everything.
If a prompt fails
The assistant passes on AtriumKit’s error, so read it first:
- “This connection wasn’t allowed to do that (Publish). Reconnect AtriumKit and allow it.” The connection doesn’t have that permission. Your client may ask you for it. If not, disconnect, connect again and tick it.
- “No “Members” in the copy… Text on the screens isn’t copy.” The word isn’t in the copy. If it shows on captured screens, change it in the editor.
- “AI translations are part of Pro and Team.” Your plan doesn’t include them.
- “Your Free plan includes 3 published demos. Upgrade or unpublish one.” Unpublish a demo or upgrade.
- “Verify your email to make and export videos on the Free plan…” Use the link AtriumKit sent you, then ask again.
- Your client says it lost access, or asks you to sign in again. The connection was disconnected, or went 90 days without use. Connect again and approve it on the AtriumKit page.
Other demo tools with MCP servers
If you already use another demo tool, start with its server. This is what each vendor’s docs said on November 5, 2026, with AtriumKit last for comparison. For 16 demo and guide tools side by side, Walnut, Reprise and HowdyGo among them, see which demo tools work with Claude, ChatGPT and Cursor.
- Storylane:
https://mcp.storylane.io/mcp, on Starter and above. Claude and ChatGPT use OAuth, and Cursor uses a client ID and secret. Tools include adding steps and publishing. - Supademo:
https://mcp.supademo.com/mcp, with OAuth. That page doesn’t say which plans include it. It says anything in the Supademo editor is available over MCP, andpublish_demomakes a version live. - Navattic:
https://app.navattic.com/api/mcp, on Startup and above. Claude and ChatGPT use OAuth, and Cursor uses a personal access token. It edits flows and publishes demos. - Arcade:
https://mcp.arcade.software/mcp, on Free, Pro and Enterprise. This server covers brand kits and AI videos. Its docs say it can’t create or edit interactive demos, and that videos can’t be edited over MCP after they’re generated. Arcade’s separate interactive demo tools read demo data, with insights on Growth and up, and can’t edit demos either. - AtriumKit:
https://app.atriumkit.com/mcp, on every plan, with OAuth or a token. It builds demos from public web pages or screenshot links, edits demo copy and video scenes, translates and publishes. It can’t edit captured screens or export videos.
Start with one read-only prompt
Connect with only Read demos, run the first prompt and see what comes back. When you’re ready, reconnect with Edit drafts and try the dry run in task 2. Create a free account if you don’t have one, and see the help article for every permission and limit.