Guide
Connect Claude to Lollipop and brief an ad without leaving the chat
Lollipop connects to Claude as a custom MCP connector. You add one address, sign in on Lollipop’s own consent screen, and pick the workspace it may work in. After that you describe the product in the chat, approve the format and the script it comes back with, and it drives the ad through to a finished cut and hands you the ad’s web address.
How do you connect it?
There is one address, https://mcp.lollipopads.ai/mcp, and it is a standard remote MCP server spoken over HTTP. Both setup paths below point at that same string.
In Claude Code, add it from your terminal with claude mcp add --transport http lollipop https://mcp.lollipopads.ai/mcp, then start a session, run /mcp, choose Lollipop and authenticate. In Claude on the web, in the desktop app or in Cowork, open Settings, then Connectors, then Add custom connector, and paste the same address. Either way your browser lands on the Lollipop consent screen and the connection is live as soon as it hands back.
Those are the paths Lollipop has walked end to end. Because the connector is ordinary MCP over HTTP, any client that speaks the protocol can in principle reach it, but the ChatGPT and Codex paths are not verified and you will not find invented menu names here for them. Wrong setup steps are worse than none.
Claude can only connect to an account that already exists.
Make your first adEvery new account starts with 4,000 credits.
What are you agreeing to on the consent screen?
It is an OAuth login, not an API key. The assistant registers itself, sends you to Lollipop to sign in, and you approve the connection on a Lollipop consent screen where you choose which workspace and which project it may work in. No secret is ever pasted anywhere.
The screen grants one fixed capability, to create and work on ads in the workspace you pick, and it is a constant in Lollipop’s own code rather than something the connecting app asks for. That distinction is doing real work. Client registration is anonymous, so an app’s self-declared name proves nothing, and the screen leads with the host your browser will be sent back to instead. You can disconnect an app later from Settings, and doing so kills the grant and every token under it.
What does it cost, and when does the meter start?
An assistant driving Lollipop over the connector spends credits from the same wallet at the same prices as the web app. There is no sandbox and no preview mode: every shot it generates is real work you have paid for.
1,000 credits is $1. Lollipop has two self-serve plans: Pro at $99/month and Scale at $249/month. Nothing about the connector changes either number.
Spending is quoted before it happens. The assistant’s first attempt to generate arrives without a confirmed figure, spends nothing, and comes back carrying the real estimate; only a second attempt that names a number at least as big goes ahead.
Read that for what it is, because the difference matters when it is your wallet. The check enforces that the model learned the price. It does not enforce that you saw it. Lollipop instructs the assistant to say the figure out loud and wait for you, and a well-behaved one does, but the mechanical guarantee is only that a price cannot be skipped.
What actually happens after you describe the product?
The instructions Lollipop hands the connected assistant start from one premise: it is the only interface you have. A tool result has been shown to the assistant and not to you, so anything you need to see it has to print, and any decision Lollipop needs it has to ask for in plain language.
So the run reads as a conversation with pauses in it. It fetches the live format catalogue and offers you a few of them in its own words. It writes a script and shows it to you word for word, because those are the words you are approving. It relays the shot list’s open questions one at a time and will not answer them on your behalf. Then it plans the visuals, generates, assembles, and warns you before the slow steps rather than going quiet for five minutes.
One optional step is worth taking first. Give Lollipop your website and it researches the brand, voice, product and positioning, before it writes anything.
Keep the link it gives you. Every ad an assistant creates has a Lollipop web address ending in /ads/ and the ad’s id, returned in the assistant’s first reply. That link is the handoff: it never expires, and it is where the canvas, the timeline and everything the assistant cannot do live.
The order those steps run in is fixed, and skipping one is refused rather than tolerated. Why it is built that way, which tools the connector exposes and which it withholds, is the other guide: what an assistant can do to an ad pipeline over MCP, step by step.
Which half of the job belongs in the chat?
Worth knowing before you start, because a model will politely attempt the wrong half and then tell you it went well.
| Format | In the Claude chat | On the ad screen |
|---|---|---|
| The script | Written, printed to you verbatim, and approved in the conversation | Readable there too, and a plan question can be answered from either side |
| Watching a shot | It cannot. It describes a take from its prompt and says it is describing | Where every judgement about how something looks is made |
| Casting by eye | It can tell you who the plan cast and why | Comparing faces and picking one |
| Your own footage | No upload path, apart from a product image pulled from a URL | Upload there, then the assistant picks the asset up |
| Timeline and captions | Out of scope, and it is told to say so rather than to try | Trimming, reordering, caption styling, music and safe zones |
| Flagged shots | It relays the link and waits | Reviewing what the quality checks held back |
What you will still do on the ad screen
- It never sees the ad. An assistant driving Lollipop over the connector cannot see images or video. Every judgement about how a shot looks belongs to you, on the ad screen, and it is instructed to describe a variant from its prompt and metadata rather than pretend to have watched it.
- You cannot hand it a file. There is no upload path through the connector. To use your own footage or product photos you upload them on the ad screen, and the assistant picks them up from there. The one exception is a product image imported from a URL.
- Finishing happens where the pixels are. A finished ad exports as an editable timeline for DaVinci Resolve, Final Cut Pro and Premiere Pro, plus a media handoff for CapCut, and none of that is reachable from a chat window.
The shortest version
Add the connector, approve the connection, then say what you sell, who it is for, and what you want them to do. If you have no view on the format, say so and let it read you the catalogue. The plainest first ask is the Yapper format, which keeps one person on camera with no cutaways, and the format library, with a real generated ad in each is worth two minutes before you brief anything.
Naming one is worth the effort. Picking a format in Lollipop is not filling in a template: each of the 21 formats ships its own scripting agent, its own scene agent and its own visual rules, and generation is bound by them.
Everything above assumes the assistant. The short account of how a video ad gets made with AI end to end is the same job without one.
Connector access is not free everywhere in this category. Creatify’s MCP server is a feature of its $99 a month Pro plan, and its API is a further separate subscription starting at $99 a month. That is worth knowing before you compare one setup guide with another.
Create the account, add the connector, brief the first ad.
Make your first adDescribe the product in a chat. Lollipop writes the script, casts the actor and renders every shot, and you can open the canvas to change any of it.