n8n UGC Video Ads: Build a Reaction-Hook Agent Workflow

Short answer: in n8n, connect an AI Agent node to the RE/UGC MCP server with the MCP Client Tool, give it a system message that describes your ad rules, and trigger it from a chat, a form or a spreadsheet of briefs. The agent searches a library of hyperrealistic AI UGC reaction clips, returns three to five hook options with previews and fetches download links for the clips you own. You can import the finished template below and have it running in about ten minutes.

If you would rather work in a chat than a canvas, see how to make UGC ads with Claude. This guide is for growth marketers and agencies who already use n8n and want UGC-style hooks without a creator brief for every variation. It covers the template, the setup, how to scale it to a batch of briefs and the guardrails worth keeping.

How this differs from AI avatar workflows

Search for n8n UGC templates and most results chain an LLM script, an image model and a video model to generate a talking avatar per ad. That works, but every render costs money and time, and you review a new face every run. A reaction-hook workflow does something narrower: it picks an existing short reaction clip as the first second of the ad and leaves the product proof to your real footage.

The two approaches are not exclusive. Many teams generate a few avatar explainers and run reaction hooks for fast variation testing. If you are deciding between them, read reaction hooks vs AI UGC.

The workflow at a glance

The MCP Client Tool gives the agent six tools: list_categories, search_clips, get_clip, get_account, get_download_link and unlock_clip. Only unlock_clip spends anything, and the template’s system message forbids calling it without an explicit yes.

Set it up in ten minutes

1. In RE/UGC, open Dashboard → AI agents and create an API key. Copy it once; it is shown only at creation.
2. In n8n, open a new workflow and use the ⋯ menu → Import from File to load the template, or add an AI Agent node and attach an MCP Client Tool sub-node.
3. In the MCP Client Tool, set Endpoint to https://reugc.com/mcp and Server Transport to HTTP Streamable.
4. Set Authentication to Bearer Auth and create a credential with your API key. Alternatively choose MCP OAuth2 and sign in with your RE/UGC account.
5. Attach a chat model credential to the model node. The template uses Anthropic; OpenAI and others work the same way.
6. Open the chat and send a brief, for example ‘Shocked reaction under 4 seconds for a budgeting app’.

The system message that keeps the agent useful

The RE/UGC server already sends the agent its workflow rules when it connects. Your system message should add what only you know: your brand, your claims and your approval policy.

You help our team pick reaction hooks for short vertical ads.
Brand: [name], [one-line product description]
Audience: [who]
Claims we can prove: [list]. Never suggest captions beyond these.
Workflow: call list_categories once, then search_clips with a category and max_duration_seconds 5.
Reply with 3-5 options: title, duration, owned or not, preview link, and one sentence on why it fits.
Only call unlock_clip after the user explicitly confirms spending 1 credit on that specific clip.
For owned clips, use get_download_link and remind the user the link expires in 5 minutes.
Also propose 3 hook captions (max 7 words) per option.

Scale it: one brief per row

The chat version is perfect for one ad at a time. For a weekly creative sprint, swap the trigger and run every brief in a spreadsheet through the same agent.

1. Create a sheet with columns: product, audience, claim, feeling, platform, status.
2. Replace the chat trigger with a Schedule Trigger (or a manual trigger) followed by Google Sheets → Get rows filtered to status = new.
3. Feed each row into the AI Agent with a prompt built from the columns, and ask for options only, with no unlocks.
4. Write the agent’s options back to the sheet (or post them to Slack) and set status = review.
5. A person picks a clip per row. A second, small workflow unlocks only the approved clip IDs and fetches the download links.

Splitting ‘suggest’ from ‘unlock’ keeps credit spending deliberate even when the suggestion step runs unattended. It also leaves an audit trail: the sheet shows who approved which clip and when.

How many briefs and variations per sprint? Start small. Three reactions crossed with two captions on the same footage is usually enough to learn something; how many ad variations to test walks through the maths, and how to scale winning ad creatives covers what to do once one wins. Scheduled runs are also a cheap way to keep fresh hooks ready before creative fatigue sets in.

From clip to finished ad

The agent hands you the hook, the captions and a fresh download link. Rendering the final MP4 is a separate step, and you have three good options:

• Manual edit: download the clip and assemble it in CapCut with your footage. See the CapCut caption guide.
• Claude Code: let Claude download the clip and cut it with ffmpeg on your machine. See Claude Code video editing.
• A render node: add an Execute Command node with ffmpeg on a self-hosted n8n, or call a hosted rendering API, using the same cut list the agent wrote.

Whichever you choose, download the clip in the same run that fetched the link. Links expire after five minutes, and the agent can always fetch a new one for a clip you own.

Troubleshooting