GitHub - nmbrthirteen/podcli: CLI/MCP/Platform to generate clips from podcast.

AI-powered podcast content studio. Transcribe episodes, find viral moments, render upload-ready Shorts with burned captions — then generate titles, descriptions, thumbnails, and a full publish-ready content package. All from your terminal.

What It Does

podcli takes a long-form podcast and turns it into a complete content operation:

Record episode
    ↓
Transcribe (Whisper, speaker detection)
    ↓
Find viral moments (Claude AI + audio energy + knowledge base)
    ↓
Render clips (9:16, captions, smart crop, normalized audio)
    ↓
Generate content package (titles, descriptions, thumbnails, SEO)    ← PodStack
    ↓
Publish with optimization checklist                                  ← PodStack
    ↓
Review performance                                                   ← PodStack

The first half is video processing — podcli's core engine. The second half is content workflow — powered by PodStack, a set of Claude Code slash commands that ship with podcli. Both halves are deeply integrated: the clip suggestion engine reads from your PodStack knowledge base, uses your title formulas and voice rules, checks the episode database for duplicates, and outputs MCP-aligned fields that flow through to export.

How It Works (From a User's Perspective)

1. Drop in your episode

./setup.sh --ui
# → http://localhost:3847

Drag your video into the Web UI, or use the CLI:

./podcli process episode.mp4 --transcript transcript.txt --top 8

2. Get clips automatically

podcli uses Claude to analyze your transcript against your show's knowledge base, finding the most viral moments. It scores each one on 4 dimensions, suggests clips with multi-cut segments (cutting out filler), and lets you toggle them on/off before rendering.

Clips come out as upload-ready Shorts: 1080x1920, 9:16 vertical, with burned-in captions, normalized audio, and your logo.

3. Generate the full content package

Open the project in Claude Code and run:

This runs the PodStack pipeline — a gstack-style workflow that gives you:

• 8-15 scored moments with timestamps, categories, and reasoning
• 8 title options per clip following your show's title spec (verified against 6 quality gates)
• Ready-to-paste descriptions with hooks, guest attribution, hashtags, SEO keywords
• Thumbnail briefs for both podcast (16:9) and shorts (9:16) formats
• Brand review that catches banned words, voice violations, and weak hooks
• Publish checklist covering pre-upload, at-publish, first-24-hours, and day 3-4 optimization

4. Publish and track

Run /publish-checklist when uploading. A week later, run /retro-episode with your YouTube Studio stats to see what worked and what to improve.

The Two Halves

Both halves share the same knowledge base (.podcli/knowledge/) — your show's brand, voice, title formulas, episode database, and style guide. Set it up once, everything stays on-brand.

Features

Video Processing

• AI clip suggestion — Claude/Codex-powered moment detection with knowledge base context, multi-cut segments, 4-dimension scoring
• Face tracking — YuNet face detection, exponential-smoothing camera, split-screen support, speaker-aware tracking with snap cooldown
• Burned-in captions — 4 styles: branded, hormozi, karaoke, subtle
• Hardware-accelerated encoding — VideoToolbox (Mac), NVENC (NVIDIA), VAAPI, CPU fallback
• Smart cropping — center crop or face tracking (handles split-screen, Riverside-style mixed layouts)
• Multi-segment clips — automatically cuts out filler, long pauses, and tangents
• Whisper transcription — auto-transcribe with speaker detection (tiny → large)
• Transcript import — paste Speaker (MM:SS), JSON, drag-drop .txt / .srt / .vtt

Content Workflow (PodStack)

• /process-transcript — extract and score best moments from any transcript
• /generate-titles — 8 titles per clip with 6-point verification checklist
• /generate-descriptions — descriptions + hashtags + SEO keywords
• /plan-thumbnails — thumbnail text + designer briefs for both formats
• /review-content — paranoid brand check (banned words, voice, title rules)
• /prep-episode — full pipeline: transcript → publish-ready package
• /publish-checklist — pre/post-publish optimization
• /retro-episode — performance analysis after publishing

Infrastructure

• Knowledge base — .md files that teach the AI your brand, voice, and style
• Asset management — register logos and videos for quick reuse
• Clip history — tracks everything to avoid duplicates
• Preset system — save named configurations per show
• MCP server — 17 tools for Claude Desktop / Claude Code integration
• Web UI — single-page flow at localhost:3847
• CLI — one-command processing: ./podcli process video.mp4 --top 5

Prerequisites

Quick Start

git clone https://github.com/nmbrthirteen/podcli.git
cd podcli
chmod +x setup.sh podcli
./setup.sh

This will:

1. Check system dependencies (Node, Python, FFmpeg)
2. Create a Python virtual environment and install packages
3. Install Node packages and build TypeScript
4. Set up PodStack slash commands and knowledge base templates
5. Create the local .podcli/ data directory
6. Launch the web UI at http://localhost:3847

Setup options

./setup.sh              # full install + launch UI
./setup.sh --install    # install only
./setup.sh --ui         # launch UI only (skip install)
./setup.sh --mcp        # print MCP config for Claude

Usage

Web UI

./setup.sh --ui
# → http://localhost:3847

1. Set video — drag-and-drop or enter a local path
2. Add transcript — drag a .txt file, paste Speaker (MM:SS) text, or auto-transcribe with Whisper
3. Generate Clips — analyzes audio energy + transcript to suggest viral moments
4. Review — toggle clips on/off, pick caption style, crop mode, logo
5. Export — batch-renders selected clips with hardware acceleration
6. Preview / Download — watch results inline, download individual clips

CLI

# Auto-transcribe + suggest top 5 clips + export
./podcli process video.mp4

# With existing transcript
./podcli process video.mp4 --transcript transcript.txt --top 5

# Full options
./podcli process video.mp4 \
  --transcript transcript.txt \
  --top 8 \
  --caption-style branded \
  --crop center \
  --logo logo.png

Presets

./podcli presets save myshow --caption-style branded --logo logo.png --top 5
./podcli presets list
./podcli process video.mp4 --preset myshow

Content Workflow (PodStack)

Open the project in Claude Code, then use slash commands:

# Full pipeline — transcript to publish-ready package
/prep-episode

# Individual steps
/process-transcript        # extract moments from a transcript
/generate-titles           # get 8 title options for a clip
/generate-descriptions     # get descriptions + hashtags
/plan-thumbnails           # get thumbnail briefs for your designer
/review-content            # brand and quality review
/publish-checklist         # pre/post-publish ops
/retro-episode             # performance analysis

Or just paste a transcript — Claude auto-detects the input and runs the right command.

Knowledge Base

The knowledge base is what makes podcli understand your show. Drop .md files into .podcli/knowledge/ and both the video engine and content workflow use them. The clip suggestion engine reads 8 of these files (prioritized by relevance), checks the episode database for duplicate avoidance, and applies your voice rules and title formulas when generating suggestions.

PodStack ships with 13 starter templates that you fill in with your show's details:

Manage via the web UI at /knowledge.html (drag & drop, inline editor) or through the knowledge_base MCP tool.

MCP Server (Claude Integration)

podcli is a Model Context Protocol server — Claude can use it as a tool to create clips through conversation.

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "podcli": {
      "command": "node",
      "args": ["/path/to/podcli/dist/index.js"],
      "env": {
        "PYTHON_PATH": "/path/to/podcli/venv/bin/python3"
      }
    }
  }
}

Claude Code:

claude mcp add podcli -- node /path/to/podcli/dist/index.js

Run ./setup.sh --mcp to get the exact config with your paths filled in.

MCP Tools

Caption Styles

Project Structure

podcli/
├── podcli                    # CLI entry point
├── setup.sh                  # one-command install & launch
├── package.json
├── CLAUDE.md                 # PodStack master config
│
├── .claude/commands/         # PodStack slash commands
│   ├── process-transcript.md
│   ├── generate-titles.md
│   ├── generate-descriptions.md
│   ├── plan-thumbnails.md
│   ├── review-content.md
│   ├── prep-episode.md
│   ├── publish-checklist.md
│   └── retro-episode.md
│
├── src/                      # TypeScript
│   ├── index.ts              # MCP server entry (stdio)
│   ├── server.ts             # MCP tool definitions
│   ├── config/paths.ts
│   ├── models/index.ts
│   ├── handlers/             # MCP tool handlers
│   ├── services/
│   │   ├── python-executor.ts
│   │   ├── file-manager.ts
│   │   ├── asset-manager.ts
│   │   ├── clips-history.ts
│   │   ├── knowledge-base.ts
│   │   └── transcript-cache.ts
│   └── ui/
│       ├── web-server.ts     # Express server + API
│       └── public/           # Frontend (React SPA)
│
├── backend/                  # Python
│   ├── main.py               # stdin/stdout JSON dispatcher
│   ├── cli.py                # CLI entry point
│   ├── presets.py
│   ├── requirements.txt
│   ├── models/               # ML model files
│   │   └── face_detection_yunet_2023mar.onnx
│   ├── services/             # Whisper, FFmpeg, captions, face tracking, etc.
│   │   ├── face_detector.py  # shared YuNet face detector
│   │   └── ...
│   └── config/
│       └── caption_styles.py
│
└── .podcli/                  # local data (gitignored)
    ├── knowledge/            # .md context files for AI (13 templates)
    ├── assets/               # registered logos, videos
    ├── cache/transcripts/    # cached transcriptions
    ├── history/              # generated clip history
    ├── output/               # rendered clips
    ├── presets/              # saved configurations
    └── working/              # temp files

Configuration

Copy .env.example to .env (setup.sh does this automatically):

Transcript Format

Speaker Name (00:00)
What they said goes here as plain text.

Another Speaker (00:45)
Their response text here.

The time offset field (default: -1s) shifts all timestamps to sync with audio.

Credits

Content workflow powered by PodStack — inspired by gstack by Garry Tan.

License

MIT