Find Viral TikTok Videos in 4 Steps (With an AI Agent)

Four steps, no dashboard

This is the agent version of TikTok research. You connect once, then work in plain language — no tabs, no CSV exports, no copying numbers into a chat and hoping the context survives.

If you would rather do it by hand, how to find viral TikToks before they blow up covers the signals without any of this machinery. This guide is for the other path.

Step 1: Connect over OAuth (about thirty seconds)

In Claude, open Settings → Connectors, choose Add custom connector, and paste:

https://api.fypnow.com/mcp

Leave the authentication fields blank. This is the step people overthink. Claude fetches a discovery document from the server, finds the authorization server, registers itself, and starts an OAuth flow — there is no API key to generate and nothing to paste into a header.

A FYPNow window opens and asks two things, both of which matter:

• Which workspace the connection acts on. It is bound to exactly one. Pick the wrong one and every tool returns a permission error with nothing in Claude explaining why.
• Read-only, or read and write. Read-only can browse and look things up but cannot analyse a new video, create an agent, search trends or generate a script — those tools are not even shown to it.

Approve, and 18 tools appear.

Step 2: Point an agent at a niche

A Viral Content Agent is a niche watcher: a name and up to three keywords. It searches TikTok daily and builds a catalogue of what actually travelled in that niche.

Two dozen niches already exist — Crypto, Fitness, Food & Drink, Reddit Stories, Real Estate and more — so for most topics you are browsing rather than building. If yours is missing:

Create a Viral Content Agent called "Home Espresso" watching the keywords espresso, coffee setup and barista.

Then ask what it found:

What are the top videos in the Crypto niche, ranked by outperformance, and what hooks do they use?

Outperformance, not views — this is the part that makes the answer useful. From the live Crypto catalogue:

The 9.7M video is the bigger number. The 2M video from a 2,194-follower account is the better signal — nothing about that account's reach explains it, so the content did. That is a format you can borrow; a large account doing large numbers only tells you they have a large audience.

Step 3: Use semantic search to find more of it

Once you have one video that works, the obvious next question is what else looks like this — and hashtags are a terrible way to answer it. Creators in the same format rarely use the same tags.

Semantic search solves this properly. Every analysed video carries a vector embedding of its content, so:

Find videos similar to this one.

returns videos with comparable structure and subject matter even when they share no words or hashtags. A "big number on screen, no narration, fast cuts" video finds other videos built the same way, regardless of topic.

One honest caveat: roughly 63% of analysed videos carry an embedding, so an empty result sometimes means "not embedded yet" rather than "nothing similar exists". The API says which, rather than silently returning nothing.

Step 4: Turn it into a script, in the same conversation

Analyse the respawnedreality video, then write me a script in that structure for my niche.

The teardown returns hook type and text, hook end time, scene count and timings, transcript with timestamps, retention signals and pattern interrupts. The script generator then drafts from the patterns your own account has already won with rather than from generic templates.

The number worth stealing is the hook end time. In the Crypto teardowns every winning hook landed between 2.1 and 3.8 seconds. If yours runs six, that is a concrete, testable fix — and you found it without opening a dashboard.

What this actually costs you in time

Be realistic about one thing. Reads are fast — browsing a niche, looking up an existing analysis and semantic search all return in well under a second.

A fresh AI teardown takes about a minute per video, because it runs a video model over real footage. Ask for hooks across five never-analysed videos and you will wait a few minutes while the assistant polls. Videos already analysed come back instantly.

So frame the request well: ask about three or four videos, not thirty. That is the difference between a ten-second answer and a five-minute one.

Why the harness matters more than the model

The reason this works is not that the assistant is clever. It is that the numbers arrive with their frame intact.

Paste "this video got 400,000 views" into a chat and the assistant cannot tell you anything useful, because the signal is the ratio to that creator's baseline and the baseline did not survive the copy-paste. Paste a transcript without scene timings and it cannot tell you the hook ran 3.8 seconds before the first cut, because it never saw the cuts.

Connected, those relationships come through. The assistant is not guessing at what virality looks like — it fetched the measurement.

Where to go next

• Connect TikTok data to Claude — the setup in more detail, plus the prompt library
• Connect TikTok data to ChatGPT — custom actions and where the MCP route stands
• The MCP server — all 18 tools and what each one costs
• How to find viral TikToks before they blow up — the signals, done manually