I Built a Local Video Processing Workstation with AI - Here's the Complete Journey

The Problem

Most video processing tools force you to:

• Upload files to the cloud (privacy concerns)
• Deal with file size limits
• Pay for premium features

I wanted a fully local, feature-rich, good-looking video processing tool. So I built ClipForge.

What is ClipForge?

ClipForge is a desktop app that handles 20+ video/audio operations locally:

• Transcode - MP4, WebM, MKV, MOV, AVI, GIF
• Visual - Crop, watermark removal, rotate, color adjust, denoise
• Speed - 0.25x to 4x, reverse, boomerang, loop, fade
• Audio - Extract, mute, volume, normalize
• Composite - Concatenate, side-by-side, picture-in-picture, overlay, subtitles

Three modes: Single operation, Stack (chain multiple ops), Batch processing.

Built with Electron, React, FFmpeg, and Zustand. Ships for Windows, macOS, and Linux.

The Tech Stack

Development Journey

Step 1: Scaffold

npm create electron-app clipforge

Electron Forge generated the boilerplate: main process, preload script, renderer with Vite.

Step 2: UI Layout

Built a 4-panel layout:

• Left: Media pool + operation library
• Center: Real-time preview canvas
• Right: Parameter inspector
• Bottom: Stack/Batch queue + logs

Dark theme with Tailwind CSS.

Step 3: FFmpeg Integration (The Hard Part)

This is the core challenge - wrapping FFmpeg's CLI into visual operations.

Architecture:

Renderer (React)
    │ invoke('process:start', request)
    ▼
Preload (IPC bridge)
    │
    ▼
Main Process (Node.js)
    │ composeArgs(request) → ffmpeg args array
    ▼
FFmpeg (child_process.spawn)
    │ progress parsing from stderr
    ▼
Events back to renderer

Example: Watermark Removal

Instead of FFmpeg's delogo filter (which has boundary restrictions - x≥1, y≥1, no edge support), I used a crop + blur + overlay approach:

case 'delogo': {
  const x = Math.max(0, Math.round(Number(p.x) || 0));
  const y = Math.max(0, Math.round(Number(p.y) || 0));
  const w = Math.max(10, Math.round(Number(p.w) || 10));
  const h = Math.max(10, Math.round(Number(p.h) || 10));
  args.push('-filter_complex',
    `[0:v]split[a][b];` +
    `[b]crop=${w}:${h}:${x}:${y},gblur=sigma=30,format=rgba,colorchannelmixer=aa=0.7[b2];` +
    `[a][b2]overlay=${x}:${y}[out]`
  );
  args.push('-map', '[out]', '-map', '0:a?');
  args.push(...videoCodec(outExt));
  break;
}

The filter graph:

• crop - extract the watermark region
• gblur - Gaussian blur (more natural than boxblur)
• colorchannelmixer=aa=0.7 - semi-transparent blend for smooth integration

Step 4: Real-time Preview

Users need to see changes immediately, not after processing completes.

Solution: Canvas-based preview simulation. Instead of running FFmpeg, read frames from the <video> element and apply operations on a <canvas>:

useEffect(() => {
  const render = () => {
    drawPreview(ctx, video, previewOps, {
      width: rect.width,
      height: rect.height
    });
  };
  render(); // immediate draw
  if (playing) {
    const loop = () => {
      render();
      raf = requestAnimationFrame(loop);
    };
    raf = requestAnimationFrame(loop);
  }
  return () => cancelAnimationFrame(raf);
}, [playing, playhead, JSON.stringify(previewOps)]);

Adjusting brightness, crop region, or rotation shows instant feedback.

Step 5: Mouse Region Selection

For watermark removal, users drag to select the area. Screen coordinates must convert to video pixel coordinates (accounting for letterbox scaling):

function screenToVideo(localX, localY, container, videoW, videoH) {
  const { scale, ox, oy } = getVideoMapping(container, videoW, videoH);
  return {
    x: Math.max(0, Math.min(Math.round((localX - ox) / scale), videoW)),
    y: Math.max(0, Math.min(Math.round((localY - oy) / scale), videoH)),
  };
}

Bug I hit: The onUp callback captured stale state from useState. Fixed by using useRef for live coordinates during drag.

Step 6: Packaging & CI/CD

Electron packaging is tricky - FFmpeg binaries can't go inside the asar archive, and Linux needs lowercase executable names.

forge.config.ts:

packagerConfig: {
  asar: {
    unpackDir: 'src/main/ffmpeg'
  },
  extraResource: ['src/main/ffmpeg'],
  executableName: 'clipforge',
}

GitHub Actions builds all three platforms in parallel:

jobs:
  build:
    strategy:
      matrix:
        os: [macos-latest, ubuntu-latest, windows-latest]
    runs-on: ${{ matrix.os }}
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
      - run: npm ci
      - run: npm run make

Push a tag → auto-build → auto-publish to GitHub Releases.

Lessons Learned

AI Coding is "Efficient Coding", Not "No Coding"

Claude Code handled tedious work (Electron packaging, FFmpeg arg mapping, IPC boilerplate), but I still needed to:

• Make architectural decisions
• Review generated code
• Debug edge cases (stale closures, boundary conditions)

Ship the Core Flow First

Got "open file → select operation → process → output" working before adding preview, batch mode, or i18n.

Packaging is the Last Minefield

Binary files, asar compression, platform-specific naming - expect to spend time here. Automate with CI early.

The Result

• 3 weeks of part-time work
• 20+ operations across 6 categories
• 3 platforms supported
• Fully local processing
• Open source: github.com/mayu888/clipforge

License: MIT + Commons Clause (free for personal use, commercial use requires authorization).

What's Next

• Drag-and-drop operations between panels
• More filter effects (LUTs, stabilization)
• Plugin system for custom operations

Built with Electron, FFmpeg, and a lot of help from AI. The future of indie development is here.

electron ffmpeg react typescript ai-coding desktop-app video-processing indie-hacker