ClipForge

Hi Fazier! I built ClipForge because I couldn't find a clip tool that didn't require uploading my source video somewhere first - relevant if you record under NDA, have slow upload speed, or just don't want your footage sitting on someone else's server. Everything runs locally on Windows: Whisper for transcription, PyAnnote for speaker labels, FFmpeg for the render. The ranking is a transparent heuristic, not a black box, so you can see why a clip scored the way it did and disagree with it. It's an unsigned soft launch (Windows shows a SmartScreen warning on first run, explained on the download page), and I'm a solo dev - happy to answer anything about the local pipeline or where it currently falls short.

Running Whisper and PyAnnote fully local is a real differentiator for NDA material - cloud clip tools are a non-starter for that. Curious how the local pipeline holds up on CPU-only machines for a 2-hour VOD: roughly what processing time should someone without a GPU expect?

The transparent scoring is the part I'd actually use. Most clip tools hand you a ranked list with zero explanation, so when clip #1 is wrong you have no idea what to adjust. Showing the reasons per candidate means I can tune for my own format (e.g. we cut product demo footage into 9:16 ads, where 'hook in the first 2 seconds' matters more than speaker changes). Two questions: can the heuristic weights be edited, and does the burned-in caption style support custom fonts/colors so the clips match brand guidelines?

A great tool for edits.

Keeping the whole clip job on the PC is the right default. Nothing uploading means faster iteration and less worry about private footage.

Keeping source footage on the user's machine is a compelling privacy advantage, and the transparent scoring helps creators decide which clips to review. A quick export-time comparison by GPU tier would make the local workflow easier to plan.

Keeping the whole pipeline (Whisper, PyAnnote, FFmpeg) local instead of routing footage through a cloud queue is the part that actually matters for the NDA/podcaster use case — most "clip your video" tools quietly assume you're fine uploading raw source, which a lot of creators aren't. Making the ranking heuristic transparent instead of a black-box score is also the right call, since it lets you correct the tool's judgment rather than just accept it. Since the optional LLM scoring uses your own API key, does that step ever send clip transcripts/metadata off-device, or does it stay opt-in per project with a clear indicator of what's leaving the machine?

ClipForge keeps source footage on the machine while Whisper transcribes on your own GPU or CPU, PyAnnote labels speakers, and FFmpeg renders the export. When product teams want to free up room in their AI budget, PZERO is at pzero.studio.

The local-only processing is the part that stands out to me - running Whisper and FFmpeg on the user's own machine means the source footage never has to be trusted to anyone else, which matters a lot for unreleased or client material. I also like that the clip ranking shows its score and the reasons behind it rather than being a black box, since that lets you overrule it before publishing. Curious whether macOS support is on the roadmap, as Windows-only is the one thing that would stop me trying it.

Its really beneficial because i need to clip long videos

Keeping transcription, speaker labels, and rendering local is a strong fit for long client footage. The visible ranking reasons are especially useful, but I’d also want the workflow to be reproducible: can ClipForge save a project with transcript edits, clip boundaries, and ranking settings, and export an SRT or EDL so the final polish can continue in Premiere or Resolve?

Running video processing entirely locally on the user machine is a huge advantage for privacy and turnaround time, especially for creators handling larger 4K footage who want to avoid heavy cloud uploads. Does it support hardware acceleration like Apple Silicon Metal or Nvidia NVENC?

The local-first workflow and transparent clip scoring are compelling, especially for NDA footage. The no-watermark trial makes it easy to validate export quality before committing; I’m curious how PyAnnote speaker labeling performs on noisy multi-speaker recordings.

The clear disclosure about the unsigned Windows build is helpful. A compact trust page with SHA-256 checksums, release notes, and the planned path to code signing could reduce SmartScreen anxiety without weakening the local-first promise. Are published checksums or signed releases on the roadmap?