MetaGPT: how an open-source agent framework reached close to half the socially active developer audience

MetaGPT (DeepWisdom) · global · 110+ creators across YouTube, Instagram, TikTok and X. AI & developer tools · YouTube · TikTok · IG · X · Q2–Q4 2025.

MetaGPT — the open-source multi-agent framework; 70k+ GitHub stars at the time of writing.

The problem with marketing to developers

Developers do not click ads and they do not read press releases. They watch someone build something. MetaGPT needed the kind of reach a consumer launch gets, from an audience that treats most marketing as noise — and it needed that reach to end in a GitHub clone, not a like.

What we did: two engines, not one funnel

The instinct on a campaign this size is to spread across as many creators as the budget allows. We did the opposite and split the video ecosystem into two jobs, each with its own creator profile, content format and success metric.

Where the AI did the work

Three places, all upstream of a human decision. It scanned creators' published history and prioritized anyone whose content actually mentioned Python, GitHub, open source or LLMs — which removed the bias toward creators the team already knew. It pulled recent performance data automatically, so sourcing used current completion-rate trends instead of a media kit. And once posts were live it classified comment sentiment at volume, which let the brand adjust content direction while the campaign was still running rather than at the wrap.

What it returned

3.1 million impressions, 138,000 engagements at a 4.4% rate, and just over 40,000 high-intent clicks into GitHub and the documentation. In a B2B technical category, a 4.4% engagement rate means the content was being consumed rather than scrolled past. The full-funnel behaviour showed up clearly: viewers found the product on a TikTok clip, went looking for the YouTube tutorial, and ended at the repository.

The completion rate is the number we care about most. Over 40% on a ten-minute technical video is not a distribution result — it is a casting result.

What we'd change

We under-invested in X early. The technical audience that ends up starring a repository is disproportionately there, and the creators who cover agent frameworks on X have small followings and unusually high intent. On later AI campaigns we started there instead of finishing there.

Find creators