He can't edit video, so he built an MCP server: 300 shorts, millions of views
wooing0306 · reddit · 2026-10-30
The author can't use CapCut or Final Cut, yet made 300 short videos with millions of views in 3 months using Claude Code and Codex. The bottleneck was revisions, not first drafts—so he built Cutfac, a Mac app with a built-in MCP server: drag over a few seconds, draw a box on what's off, type the change, and Claude Code/Codex does the edit.
Key MCP design decisions:
- listopenfeedback: the agent gets the exact clip, a red-boxed frame, and a pre-edit copy of the video—no more guessing which caption you meant.
- Each video is a plain folder with an item.yaml; the agent edits files directly and renders via local ffmpeg with a render tool.
- reviewmedia was the biggest win: the agent must inspect a contact sheet plus a strip of 6 frames around every cut (where half-finished transitions show up), dim areas covered by TikTok/Reels UI on 9:16, then name its 3 worst problems with timecodes and fix them—solving agents saying "done" without looking.
- savevariant keeps 3 versions playing side by side; keep one or undo.
19 tools total; init writes .mcp.json and registers Codex. The app deliberately has no manual editing tools; it runs on your own Claude/ChatGPT plan and never touches your footage. Open source (AGPL-3.0), free if self-built, signed app $29 one-time after 3 videos, Apple Silicon only. Open question: hand the agent frames up front, or let it pull frames itself?
Related event: Creator used Claude Code to make 300 viral short videos in 3 months(2 posts)→
More from coding & agent
- Claude Code VSCode extension reportedly backdoored via OpenVSX, not related to Cursor — jedisct1 · 2026-10-30
- AWS's EasyProver solves 669/672 Putnam problems, plus new harder benchmark Aegis — amazonwebservices · 2026-10-30
- Matt Pocock has his agent build a custom voice assistant app for Pixel Watch — mattpocockuk · 2026-10-30
- Genomics dev of 10+ years: identity crisis, then embracing agent-era coding — jmschreiber91 · 2026-10-30
- HF benchmarks models on catching reward hacking — Liquid AI's LFM decision model leads — yb2698 · 2026-10-30
- LLM Wrappers vs Harnesses: why the system around the model is where value lives — MartinGTobias · 2026-10-30