Hypit

Hypit is an open-source AI video cloning tool from Shenzhen-based Kaxi Yongliu Technology that gives coding agents like Claude Code and Codex a language for describing video (SVML) plus a compile-and-run runtime, letting them batch-replicate winning short-video formats into hundreds of variants in one command.

Core Capabilities

• Video cloning: drop in a reference video and the agent reverse-engineers the complete workflow, not just a one-off render
• Anchored to words, not seconds: subtitles/B-roll/effects bind to individual words in the script, so changing lines, languages, or presenters re-lays-out automatically — a genuine mechanism innovation
• Batch variants: one command produces 10/50/100+ hook variants, replicating viral structures at scale
• SVML video DSL: five layers (script/assets/subtitles/voiceover/compositing), described as "HTML for video"
• Model freedom + BYOK: compatible with Claude Code/Codex, bring your own model keys, free core code with no watermark

Use Cases Built for developers and growth marketing teams using Claude Code/Codex who have validated a short-video structure and need to mass-produce multi-language/multi-placement variants; also for technical teams researching "agent-driven video pipelines." Pure creators should look elsewhere.

Unique Advantages Compared with closed-source talking-head/ad tools like Arcads and Higgsfield, Hypit clones the "workflow" rather than a one-off video, is open and controllable, and lets you swap models. But official examples are only 18–26 seconds, and long-video capability is unproven.

Editorial Review A real open-source project, not a shell: 11,204 GitHub stars, 357 paid subscriptions, and Stripe-verified MRR of $5,734. But the marketing runs ahead of the evidence: the brand wall lists Google/Meta and 14 companies total with no corroboration, while primary interviews show initial customers were just two small AI startups, Cluely and Stella; the "100M views" figure has no source; the license is NOASSERTION (modified Apache 2.0 with commercial restrictions), inconsistent with the "Open Source" claim; and costs cannot be estimated upfront — Issue #270 reports a user burning 10 million tokens in the analysis phase with zero videos produced. Best for developer/agent users; start with cheaper models and set budget caps.