Here is the shape of a normal Tuesday. The creative team drops six videos in a shared folder. Somebody decides the budget, the countries and the end date in a two-line Slack message. And then somebody else spends the next hour and a half in Ads Manager turning those two lines into a campaign, three ad sets, six ads, correct crops, a naming convention and a UTM string.
Nothing in that hour and a half is a decision. It is transcription. This walkthrough is about handing the transcription over.
The setup, once
Connect an assistant to your uplads workspace. Claude and ChatGPT sign in directly from their connector settings, with no key to copy; Claude Code, Cursor, VS Code and Codex use a token you generate in uplads. The full walkthrough is in the help article on connecting an AI assistant, and the capability list is on the MCP page.
One-time job. Everything below is what the week looks like afterwards.
The brief
The message that starts it should read like something you would send a colleague:
Launch our summer sale on Meta. The videos are in this Dropbox folder: [link]. 50 € a day, Germany and Austria, running until 31 August, linking to aurorabeauty.com/summer-sale. Put it in the existing Prospecting campaign.
Five things earn their place in that message, and they are the same five that would stop a colleague from replying with a question:
- A folder link, not a file list. A Dropbox or Drive folder link imports every image and video in one go, and the assistant groups the aspect-ratio variants into multi-placement ads on its own. Naming files by hand is the slow path.
- Numbers, not adjectives. "50 € a day" and "until 31 August" produce a correct ad set. "A modest budget for a few weeks" produces a follow-up question or a guess.
- The geography. Countries, regions or cities with a radius. The assistant looks up the real location ids rather than inventing them.
- The destination. The landing page, and any UTM convention you want on it.
- Which structure to reuse. "Put it in the existing Prospecting campaign" saves it from creating a new one, which is usually what you want.
Write the brief in the language the ads should run in, unless you say otherwise. The copy comes out in that language.
What happens next
The assistant works through the job in the open, one tool call at a time, so you can watch it and interrupt:
What the assistant does with the brief
- 1
Reads the account
Ad accounts, the existing campaigns and ad sets you referred to, your creative library and the pixels on the account. It works out what already exists before it creates anything.
- 2
Imports the creatives
Every file behind the folder link, straight into your creative library. Files you already uploaded are matched rather than duplicated.
- 3
Writes the copy
Primary texts, headlines and descriptions, per creative where the creatives differ enough to deserve their own angle.
- 4
Builds the structure
Campaign and ad sets with objective, budget, runtime, the pixel event to optimise for, and targeting resolved to real ids. All of it created paused, holding no ads, so nothing can spend while you read.
- 5
Stages the launch
Everything lands in the uplads launch wizard as a draft, with a banner saying an assistant prepared it and any setup notes it could not configure through a tool.
What comes back to you is a link.
The review screen is the job
Open the link and you are in the ordinary uplads launch wizard, on the creatives step, with everything already selected. This is the part people underestimate before they have used it: because it is the normal wizard rather than a special "AI result" view, everything works exactly as it always did. Swap a creative. Rewrite the second headline. Add an ad set it did not know about. Change the ad status so the ads go live paused.
Then press Launch, and it is an ordinary uplads launch from there - background processing, per-ad failure isolation, live progress.
The reason the flow ends here rather than at a live ad is not caution for its own sake. It is that the assistant is genuinely good at the mechanical half and genuinely not the one who should decide whether the third angle is worth spending on. Splitting the job at the review screen puts each side where it is strong.
Try the connector on one real brief
The uplads MCP connector is in beta and included in every plan. Connect Claude or ChatGPT, hand it one real weekly brief, and compare the clock against the current Tuesday.
When you want it to launch as well
There is a version of this where you are not at a laptop at all: brief it from your phone on the way somewhere, come back to live ads.
Autonomous launching is a per-workspace switch, off until you turn it on, and it takes a daily budget ceiling. If the ad sets in the draft add up to more than that per day, the assistant refuses and hands you the review link instead of launching. Every autonomous launch is written to your audit log.
Worth being clear-eyed about: this is the only setting in uplads that lets something spend money without a human clicking Launch. The ceiling is what makes it a bounded decision rather than a leap. Most teams we talk to run staged-draft mode for the first month and only then decide whether the review step was ever catching anything.
Where it fits in an agency week
The pattern that has worked best is not "the assistant runs the account". It is narrower and more useful than that:
- Monday brief, Tuesday build. The brief message becomes the draft the moment it is written, instead of joining a queue.
- Client-by-client repetition. The same brief shape across eight accounts is exactly the work that is expensive for a human and free for an assistant.
- The odd hours. A brief that lands at 19 does not have to wait for tomorrow to become a reviewable draft.
- Everything after the launch stays human. Reading the numbers, killing the losers, deciding what to test next - the AI Optimizer proposes, you approve, and none of that changes because a connector exists.
For the wider view of what MCP is and how to judge a connector before you plug one into a client account, read MCP for advertisers.